Patent Grant
8237819
Reference is made to commonly assigned, co-pending U.S. Patent Applications:
Ser. No. 12/512,070 by Ptucha et al. filed of even date herewith entitled “Method for Matching Artistic Attributes of a Template and Secondary Images to a Primary Image”;
Ser. No. 12/512,056 by Bogart et al. filed of even date herewith entitled “Artistic Digital Template for Image Display”;
Ser. No. 12/512,061 by Bogart et al. filed of even date herewith entitled “Method for Producing Artistic Image Template Designs”;
Ser. No. 12/512,064 by Whitby et al. filed of even date herewith entitled “Apparatus for Generating Artistic Image Template Designs”;
Ser. No. 12/512,060 by Whitby et al. filed of even date herewith entitled “Method of Generating Artistic Template Designs”;
Ser. No. 12/512,063 by Ptucha et al. filed of even date herewith entitled “System for Coordinating User Images In An Artistic Design”;
Ser. No. 12/512,057 by Bogart et al, filed of even date herewith entitled “Coordinating User Images in an Artistic Design”;
Ser. No. 12/512,054 by Shkurko et al. filed of even date herewith entitled “Image Capture Device with Artistic Template Design “;
Ser. No. 12/512,067 by Shkurko et al. filed of even date herewith entitled “System For Matching Artistic Attributes Of Secondary Image And Template To A Primary Image”;
Ser. No. 12/512,077 by Ptucha et al. filed of even date herewith entitled “Artistic Digital Template for Image Display”;
Ser. No. 12/512,117 by Ptucha et al. filed of even date herewith entitled “Method of Making an Artistic Digital Template for Image Display”;
Ser. No. 12/512,129 by Ptucha et al. filed of even date herewith entitled “Context Coordination for an Artistic Digital Template for Image Display”;
Ser. No. 12/512,144 by Ptucha et al. filed of even date herewith entitled “Processing Digital Templates for Image Display”; the disclosures of which are incorporated herein by reference in their entireties.
The present invention relates to computer systems, methods, and software programs for use in making digital image products and digital image displays. In particular, the present methods and apparatuses automatically search for, find, evaluate, and arrange digital images in a digital image template according to programmed artistic designs.
Personalized image collages, clothing, albums and other image enhanced items are becoming increasingly more accessible as digital printing technologies improve. However, as personalized image bearing products have become more accessible, consumers have become more discriminating. In particular, consumers now seek methods and systems that produce customized image products in a more convenient, faster, seamless, automatic, and integrated manner. While becoming somewhat more common, many items for displaying and/or including embedded customized images are still considered novelties. Methods for recognizing image contents that fulfill a predescribed aesthetic appearance often fall short of expectations. For example, many products with customizable embedded images include photos of people. For this type of product, it would be desirable to identify images that satisfy preselected artistic criteria and/or image attributes, such as number of persons pictured, who is pictured, what zoom ratio, temporal aspects, clothing, background, season, facial expressions, hue, colorfulness, texture, and sharpness, etc. Because some artistic aesthetic elements work better for certain product formats, it would also be desirable if multiple image attributes including aesthetic criteria could be evaluated in parallel for a number of images, and the images with the highest fitness score be automatically determined by computer algorithm. Furthermore, customer satisfaction can be improved if the customer can select a favorite image, or several favorite images, to be included in an image enhanced product. Attributes from this favorite image can be extracted in real time from the image itself, read from a file associated with the image, such as a metadata file, a DPOF file, or a file stored by a user, or it can be generated in some other fashion, or even input directly by a user, and then be used in coordinating surrounding image and template contents. Conversely, attributes of a selected image product template can be used in coordinating which image or images from a user collection can be placed within the template. Furthermore, attributes of a context or of an environment where the image will be displayed can be input to a computer system and referenced to select appropriate templates and images, or any combination of these algorithms can be integrated in a computing system or method.
The problem in the prior art is the time consuming process requiring users to manually search through many images and compose their own artistic creations, for which they may not be capable.
An embodiment of the present invention includes a method of operating a digital image capture device wherein a plurality of digital image template files are stored in a memory of the capture device, each of the digital image template files include features of a digital image template such as a number, location, and shape of openings in the digital image template. Image attribute requirements are also stored in association with each of the openings for screening images to be disposed in an opening. The templates are displayed to a user for selecting a template. Images are captured and stored and each of the images include image attribute data. Stored images are automatically searched to identify images that may be automatically digitally placed into openings in the digital template. The image attributes of found images are those which satisfy image attribute requirements of the openings. A file containing a digital image template together with the images digitally disposed therein is saved in the capture device for future use such as printing the file or transmitting the file over a network.
Another embodiment of the present invention includes generating a digital image display product by operation of a digital camera. A plurality of digital image template files is stored in the camera memory, each of the files defining digital image template features such as a number, location, and shape of openings in the template. The openings each have an image attribute requirement stored in associated with it. The openings are used for placing a digital image. A camera screen is used to display the templates which can then be selected by a camera user. The camera user captures and stores images in the digital camera, and image attributes are stored with them. The images are placed into the openings of a template selected by the camera user. By operation of the present invention the camera automatically searches other images stored in the camera for identifying other images that may be digitally placed into unused openings in the digital image template. Completed digital image templates together with the images digitally placed therein are stored in the camera.
Therefore, one of the advantages of the present invention is automating the tedious steps for creating an image enhanceable product by a computer system that operates upon a database of digital images using artistic stored image attribute requirements stored as recipe segments.
These, and other, aspects and objects of the present invention will be better appreciated and understood when considered in conjunction with the following description and the accompanying drawings. It should be understood, however, that the following description, while indicating preferred embodiments of the present invention and numerous specific details thereof, is given by way of illustration and not of limitation. Many changes and modifications may be made within the scope of the present invention without departing from the spirit thereof, and the invention includes all such modifications.
The source of content or program data files 24 can include any form of electronic, optical, or magnetic storage such as optical discs, storage discs, diskettes, flash drives, etc., or other circuit or system that can supply digital data to processor 34 from which processor 34 can load software, template designs and recipes, derived and recorded metadata, image files, image attributes and required image attributes or derive images and image metadata for use in automatically forming an image enhanced item. In this regard, the content and program data files can comprise, for example and without limitation, software applications, a still image data base, image sequences, a video data base, graphics, and computer generated images, image attribute information associated with still, video, or graphic images, and any other data necessary for practicing embodiments of the present invention as described herein. Source of content data files 24 can optionally include devices to capture images to create content data for use in content data files by use of capture devices located at electronic computer system 20 and/or can obtain content data files that have been prepared by or using other devices or image enhancement and editing software. In the embodiment of
Sensors 38 are optional for particular embodiments of the present invention and can include light sensors, biometric sensors and other sensors known in the art that can be used to detect conditions in the environment of system 26 and to convert this information into a form that can be used by processor 34 of system 26. Sensors 38 can also include one or more cameras, video sensors, scanners, microphones, PDAs, palm tops, laptops that are adapted to capture images and can be coupled to processor 34 directly by cable or by removing portable memory 39 from these devices and/or computer systems and coupling the portable memory to slot 46. Sensors 38 can also include biometric or other sensors for measuring involuntary physical and mental reactions. Such sensors including, but not limited to, voice inflection, body movement, eye movement, pupil dilation, body temperature, and p4000 wave sensors.
Memory and storage 40 can include conventional digital memory devices including solid state, magnetic, optical or other data storage devices, as mentioned above. Memory 40 can be fixed within system 26 or it can be removable and portable. In the embodiment of
In the embodiment shown in
Communication system 54 can comprise for example, one or more optical, radio frequency or other transducer circuits or other systems that convert data into a form that can be conveyed to a remote device such as remote memory system 52 or remote display 56 using an optical signal, radio frequency signal or other form of signal. Communication system 54 can also be used to receive a digital image and other data, as exemplified above, from a host or server computer or network (not shown), a remote memory system 52 or a remote input 58. Communication system 54 provides processor 34 with information and instructions from signals received thereby. Typically, communication system 54 will be adapted to communicate with the remote memory system 52 by way of a communication network such as a conventional telecommunication or data transfer network such as the internet, a cellular, peer-to-peer or other form of mobile telecommunication network, a local communication network such as wired or wireless local area network or any other conventional wired or wireless data transfer system.
User input system 68 provides a way for a user of system 26 to provide instructions to processor 34, such instructions comprising automated software algorithms of particular embodiments of the present invention that automatically generate artistic coordinated image displays according to template recipes. This software also allows a user to make a designation of content data files, such as selecting image templates and designating primary images, to be used in automatically generating an image enhanced output product according to an embodiment of the present invention and to select an output form for the output product. User controls 68a, 68b or 58a, 58b in user input system 68, 58, respectively, can also be used for a variety of other purposes including, but not limited to, allowing a user to arrange, organize and edit content data files, such as coordinated image displays and image templates, to be incorporated into the image enhanced output product, for example, by incorporating image editing software in computer system 26 which can be used to override automated image enhanced output products generated by computer system 26, as described below in certain embodiments of the present invention, to provide information about the user or audience, to provide annotation data such as voice and text data, to identify characters in the content data files, and to perform such other interactions with system 26 as will be described later.
In this regard user input system 68 can comprise any form of device capable of receiving an input from a user and converting this input into a form that can be used by processor 34. For example, user input system 68 can comprise a touch screen input 66, a touch pad input, a 4-way switch, a 6-way switch, an 8-way switch, a stylus system, a trackball system, a joystick system, a voice recognition system, a gesture recognition system, a keyboard 68a, mouse 68b, a remote control or other such systems. In the embodiment shown in
A simplified block diagram of a camera 76 of the present invention is shown in
When a camera is purchased, it is typically provided with a software application (located on a disc 140) for running on a user's host computer system 26 that enables the user to download applications through interface 160 into camera memory 156 or 157. The user can also select, through the software application, various program features of the application using buttons 155 to navigate through various menu features shown on display screen 154 which provides interactive feedback through a camera GUI, for example, as is well known in the art. All of this information can alternatively be preloaded on the camera during manufacturing or downloaded, via the web or a memory card reader 46 on the host computer system 26, to the removable memory card 157, which can be subsequently inserted into the camera. Typically, keyword descriptors accompany the information to enable easy access by the camera user.
After placing the memory card 36 in the camera (or disconnecting the camera from the host computer system 26), the user can operate the camera 76 to take numerous pictures, which are stored either in the internal memory 156 or in the memory card 157 (or in both). After taking pictures, the user reviews the images on the LCD screen 154, using the buttons 155 to scroll through the images. The user can then select the desired template recipes and run the template recipe algorithms as will be described below in more detail. For example, user options can include the user selection of templates and recipes, the user selection of images and image manipulation software, and the user selection of template backgrounds and text.
These applications and options are stored in and accessed from the camera memory 156, memory card 157 and, for example, are assembled via menu options and displayed on the LCD screen 154. Selections among these applications and options are made, for example, by reference to menus, keyword descriptors, and actuation of the user buttons 155. Completed selections can be stored in the internal memory 156 or the memory card 157. The camera display screen is used to monitor the look of a coordinated image template design in progress or a completed template design.
For printing, template composition, etc. the user can store print size and quality information (e.g., thermal or ink jet) of the images to be printed using standard Digital Print Order Format, or DPOF metadata tags. For example, the user might choose one “standard” (4″×6″ size) image of 2 different templates and 2 standard size images plus one “enlargement” (8″×11″ size) template image of their “favorite” vacation image together with an automatically generated artistically coordinated thematic template. The template prints can include user selected text on the template body, “creative” prints, using one of the creative backgrounds, e.g. selected on the host computer and downloaded to the camera, along with optional text input. In this case, the “favorite” image might be surrounded with one of the templates and captions provided via the host computer system 26
The user can take the image memory card 157 containing completed template files to a service provider that carries corresponding image enhanceable products that are similar to the template design. Alternately, a user can place the card 157 in a slot or memory card interface 46 of a “walk-up kiosk”. The kiosk operates as described previously with respect to
Alternately, the camera could incorporate or be connected to a wired or wireless modem, such as through communications interface 161. In this case, the completed template could be transmitted to a service provider or to a printer.
If the camera includes a transmitter, e.g., a cellular connection in the communications interface 161, the camera could include and initiate a “send” command that the user would enable after generating a completed template. This command could automatically send the completed template to the appropriate user's e-mail accounts through a network coupled wirelessly to the camera using the appropriate communications protocol (FTP, mailto, etc.). Alternately, the camera can be placed in a docking unit (not shown) containing the modem for communicating over a wired network connection. Alternately, the memory card 36 could be removed from the camera and inserted in a kiosk for operating template algorithms stored there.
Instead of, or in addition to, the user may choose to transmit one or more completed templates to a network connected “electronic photo album” account (e.g. an image database) located on a server, which could be maintained by a service provider or, alternatively, could be maintained on the user's home computer in an image database. Multiple templates may be downloaded from the host computer system 26 or from a network connected server to internal memory 156 or memory card 157. Alternatively, a low resolution version of the templates desired by the user can be stored in the camera 76, for previewing the final composite image on the LCD screen. A high resolution version of the template can be stored at the service provider or at the user's home computer to print the final composite image.
As is illustrated in
Output system 28 (
In certain embodiments, the source of content data files 24, user input system 68 and output system 28 can share components. Processor 34 operates system 26 based upon signals from user input system 58, 68, sensors 38, memory 40 and communication system 54. Processor 34 can include, but is not limited to, a programmable digital computer, a programmable microprocessor, a programmable logic processor, a series of electronic circuits, a series of electronic circuits reduced to the form of an integrated circuit chip, or a series of discrete chip components.
As used herein, an embodiment of an image enhanceable item 300, such as the example frame shown in
In addition, other embodiments of an image enhanceable item are also contemplated. Embodiments of the present invention also include electronic displays such as picture frames, digital store front displays, digital projectors, score boards such as in a stadium, or elsewhere, billboards, and interactive displays. Such displays have spatial and temporal components which are described below. The temporal component can be used to display transient information, advertising, artistic content, and can be saved for later display.
Thus, the image enhanceable item 300 as shown in
It is the intent of particular embodiments of the present invention to set forth a language, method, program, or mechanism for an artist to describe his/her aesthetic intent for each template window or template opening and for the image display overall, and for this intent to be fulfilled in an automatic manner on images in real time using a computer system such as illustrated in
As another example of the operation of an embodiment of the present invention, with a template containing two window openings, we have left and right required image attributes or left and right recipes or left and right formulas. The recipe for the opening on the left may dictate that we have a colorful non-person main subject image which fills the opening at a 120% zoom. The recipe for the opening on the right may dictate that we have a landscape image with two people; both facial poses are directed slightly to the left, each face having a 20% fill factor and both faces be offset to the left by 20%. In order to fulfill the recipes each of a plurality of computer system accessible images is evaluated, via programmed computation as described below, based on a fitness measure that calculates an image's compatibility with an opening's required image attributes, for each image with respect to satisfying the recipes for the virtual frame openings. Thus, some images might better satisfy a template opening recipe than another image as indicated by a higher score, evaluation, or compatibility calculation. The images with the highest score, automatically obtained by executing the programming, for each window opening, as against a number of stored images, is automatically chosen to fill each window opening without requiring user intervention in most instances. For most cases, the user merely needs to provide a number of digital images to the computer system, e.g. attaching a flash drive at slot 46 that contains some or all of the user's image files and, in certain, but not all, embodiments of the present invention, the user also selects a favorite image or images which is then used by the computer system as a basis to automatically select remaining images to fill out a template display and/or to automatically select a template. The user may also provide a network server location to the computer system where some or all the user's images are located. The network connected computer system will then retrieve these images for use in generating a completed image display template.
Alternatively, the scoring computation can be programmed such that an optimum arrangement of images in the window openings is one that generates an overall highest score for the entire template. An arrangement that generates the highest overall score is chosen. Identifiable or measureable characteristics that might be included to affect score calculations include, but are not limited to, identity of main subject, main subject size, number of faces, face size, facial pose, facial expression, age/gender of face, dominant color, texture analysis, object match, and time stamp. Time stamps can be effective to search for and find images captured during a particular season. These various characteristics of images are referred to herein as metadata or metadata tags, and are described in more detail below. Metadata can be recorded in association with digital images as derived (extracted) metadata, or as static (recorded) metadata. Typically, users can provide recorded metadata directly by operation of a camera or by later opening image files on a computer system and providing recorded metadata for storage at that time. Almost all digital cameras associate and store metadata with image files captured by the camera, such as time stamps, focal settings, file names, GPS location data, etc. Extracted metadata is obtained by processing image data files, either by a camera's computer system or other external hand held, or workstation, kiosk, or laptop computer system, using programmed algorithms for extracting characteristics in the image. A reference which explains computerized facial age estimation of persons in digital images is Lobo and Kwon, “Automatic Feature Detection and Age Classification of Human Faces in Digital Images”, U.S. Pat. No. 5,781,650. A reference that discusses computer executed facial expression determination is described in Yin, Loi, Xiong, “Facial Expression Analysis Based on Enhanced Texture and Topographical Structure”, Conference Proceedings—IEEE International Conference on Systems, Man and Cybemeteics, v1, P 586-591, 2004. A reference that describes digital gender estimation is Tivive, Bouzerdoun, “A Gender Recognition System Using Shunting Inhibitory Convolutional Neural Networks”, IEEE International Conference on Neural Networks—Conference Proceedings, p 5336-5341, 2006. These three references are incorporated herein by reference in their entirety. “Recipes” are discussed below with respect to preferred embodiments of the present invention.
A majority of keepsake photographic memories contain pictures of people, and, as such, people are often the main subjects in images and, therefore, often may be critical in fulfilling recipe requests. Using computer methods described in U.S. Pat. No. 7,508,961B2 by Chen, Nicponski, and Ray, or P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” in Computer Vision and Pattern Recognition, 2001, Proceedings of the 2001 IEEE Computer Society Conference, 2001, pp. I-511-I-518 vol. 1, or Schnieiderman, H. Schneiderman, “Feature-centric evaluation for efficient cascaded object detection,” in Computer Vision and Pattern Recognition, 2004. Proceedings of the 2004 IEEE Computer Society Conference, 2004, pp. II-29-II-36 Vol. 2, the size and location of each face can be found within each image. These three documents are incorporated by reference herein in their entirety. Viola utilizes a training set of positive face and negative non-face images. Then, simple Haar-like wavelet weak classifier features are computed on all positive and negative training images. While no single Haar-like feature can classify a region as face or non-face, groupings of many features form a strong classifier that can indeed be used to determine if a region is a face or not. This classification can work using a specific size window. This window is slid across and down all pixels in the image looking for faces. Then the window is enlarged and larger faces are scanned in the image. The process repeats until all faces of all sizes are found in the image. Because this process can be quite compute intensive, optimizations such as an integral image and cascades of weak classifiers makes the process work faster. Not only will this process find all faces in the image, it will return the size of each face. Similar techniques can be used to find the location and size of any non-deformable object such as people, cakes, balloons, graduation caps, cars, boats, and other desirable objects.
Once a face is found, neural networks, support vector machines, or similar classifying means can be trained to locate specific features such as eyes, nose, and mouth; and then corners of eyes, eye brows, chin, and edge of cheeks can be found using geometric rules based upon anthropometric constraints such as those described in DeMenthon, Daniel F, Davis, Larry S., “Model Based Pose in 25 Lines of Code”, Proceedings from the Image Understanding Workshop, 1992. Active shape models as described by Cootes, T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham in “Active shape models—their training and application,” Computer Vision and Image Understanding, vol. 61, pp. 38-59, 1995, can be used to localize all facial features such as eyes, nose, lips, face outline, and eyebrows. These two publications are incorporated by reference herein in their entirety. Once all features are found, it is possible to determine if eyes/mouth are open, or if the expression is happy, sad, scared, serious, neutral, or if the person has a nice smile. Determining pose uses similar extracted features. A. Savakis, M. Erhard, J. Schimmel, and J. Hnatow in “A multi-camera system for real-time pose estimation,” Bellingham Wash., 2007, developed a geometric model that adhered to anthropometric constraints. This document is incorporated by reference herein in its entirety. With pose and expression information stored as metadata for each image, the present invention can give the artist more functionality to define increasingly nuanced image enhanced products by allowing the artist to specify in a window recipe, for example, a person looking to the left with a serious demeanor.
In many instances there are no people depicted in an image, but there is a main subject that is not a person or that does not contain a recognizable face. A main subject detection algorithm, such as the one described in U.S. patent application Ser. No. 09/223,860, filed Dec. 31, 1998, which is incorporated herein by reference in its entirety, can be used to search for and find an image to fill each opening in an aesthetic manner according to stored recipes. Such a recipe might require a particular object to be present in a an image which is searchable by virtue of that object being identified in the metadata associated with an image. The metadata for images in a data base of images is searched and those images whose metadata satisfies the search are found and evaluated according to programmed evaluation procedures which are typically defined as calculations to provide a compatibility index as between a found image and a template opening's required image attributes.
Other exemplary embodiments of such algorithms involve segmenting the image into a few regions of homogeneous properties such as color and texture. Region segments can be further grouped into larger regions based on similarity measures. Regions are algorithmically evaluated for their saliency using two independent yet complementary types of saliency features—structural saliency features and semantic saliency features. The structural saliency features are determined by measureable characteristics such as location, size, shape and symmetry of each region in an image. The semantic saliency features are based upon previous knowledge of known objects/regions in an image which are likely to be part of foreground (for example, statues, buildings, people) or background (for example, sky, grass), using color, brightness, and texture measurements. For example, identifying key features such as flesh, face, sky, grass, and other green vegetation by algorithmic processing are well characterized in the literature. The data for both semantic and structural types can be integrated via a Bayes net as described by Russell and Norvig, “Artificial Intelligence—A Modern Approach,” 2nd Edition, Prentice Hall, 2003, to yield the final location of the main subject. This document is incorporated by reference herein in its entirety. Such a Bayes net combines the prior semantic probability knowledge of regions, along with current structural saliency features into a statistical probability tree to compute the specific probability of an object/region being classified as main subject or background. The main subject detection algorithm provides the location of a main subject and the size of the subject as well.
Using recipes, an artist may design and store in the computer system memory creative required image attributes for a number of template openings defining an image template recipe, thereby creating several aesthetic or practical looks for a single image enhanced item. A retail store that provides the tools described herein for customer use and purchase can choose to roll out new looks for each season, holiday, or other occasions. A store that provides these product services can license recipes to or from famous artists or celebrities. An aesthetic look can also be achieved by programmed computer system analysis of a user's images or from a user's selection of a primary or favorite image or images, determining a theme or pre-determined look from image to image, and then picking a template that is complementary to that theme or look. A preferred embodiment of the present invention contemplates an initial step of characterizing images in order to select a template based on those characterizations, and also contemplates an initial selection of a template in order to identify appropriate images for inclusion. In each case, recorded image metadata is used, or is generated by extraction and stored with the images, and is compared with metadata stored with an image template to evaluate, via an evaluation procedure, compatibility between them. A preferred embodiment of the present invention extracts characteristics of a user's designated favorite image or images and incorporates those characteristics in window recipes for remaining openings in a template.
Another embodiment of the present invention includes computer system analysis of digital images for information content such as number of people and/or their expressions and/or color content, then the images are computationally evaluated against selected aesthetic intents as defined by a template or window recipe, where such artistic or aesthetic intents include specific design layouts, each layout being created in advance by an artist or someone skilled in the art of making an image layout pleasing or enticing. For a tangible consumer product, such as a collage, frame, or photo book, the intent may include making something pleasing, creative, fun, inspirational, romantic, or to invoke past memories. For transient displays, such as the electronic displays described above, the intent can be the same as for image enhanceable consumer products, such as digital picture frames or cameras, but may be quite different for commercial applications using electronic displays of various sizes. The following are examples of commercial applications: a digital storefront may be continually updating its product line in an attractive and enticing manner, an embodiment of the present invention may include a large or small digital display in a storefront window. An alternative embodiment may include a user interactive display. A news billboard intent may be to group images by subject and relevance. A sports venue billboard or scoreboard may include displaying pictures that work well together gathered from a continual scanning and capturing of images of the attending crowd or other scenes in or near the sports venue. An interactive display can have a mix of live video frames and pre-existing digital libraries and update its content based upon user input, where the user input can be received via a traditional mouse and keyboard, or may be haptic interfaces, or the display may update its content automatically by detecting a user's facial expression and user activity, and selecting those kinds and types of images for display that elicit a particular electronically detected response. The aesthetic intent that scores highest, or works best for that particular image set is chosen automatically for display or inclusion in an image enhanceable product. Scoring is symmetric in that a template can be selected in advance and the highest scoring images are included, or one or more images can be selected in advance, and then the highest scoring template is selected based on its aesthetic compatibility with the selected images. Alternatively, when creating image enhanceable consumer products, the top n highest scoring images/templates/recipes can be presented to the customer on a computer system display, who in turn picks the image/template/recipe of his/her choice and communicates such a choice to the computer system via the user input controls.
Given an aesthetic look defined by a recipe, filling each window in an image enhanceable item can be done in at least two ways. If there are n windows and n images, a preferred embodiment of the method of the present invention can include computing all possible combinations of all images contained in an image data base in each window. If there are n windows in a virtual image template and m images in an image database, where m>n, a preferred embodiment of the present invention could have many variations to score. In the latter condition, a preferred embodiment of the invention is to place the image with the highest score into each window. In this example, each image is scored by programmed computation against each window recipe individually, and the image with the highest compatibility score for a particular window gets assigned or selected to that particular window. If an image is calculated to have a highest score that is identical in greater than one window, the window with the higher weight is considered more important and will be used to display the high scoring image. Template window weighting is described in more detail below where priority of template windows correspond to the numbering of the windows in a template. If multiple images are scored equivalently for a single window opening, metadata fields that record the number of times an image was accessed can be used to break a tie. To resolve ties, an algorithm that randomly selects an image from the group can be used. The use of weighted values for images and for template openings is described in more detail below.
Another common template recipe generation and fulfillment technique is image splitting. Image splitting is the process of taking a single customer image, and spanning that one image over 2 or more window openings. For example, the program can receive as input a single picture of a bride and groom. An application of a particular recipe will result in the bride being placed in the left window with the groom in the right window. An artist will be able to define and store recipe requirements for the size, location, and spacing of two people in a single photograph. The customer images in a product order can be analyzed using the aforementioned methods. An image with a highest score can be chosen to fulfill the product intent. If no images score above a defined threshold, the algorithm can be designed so that a different recipe is automatically chosen and the process repeats itself with the new recipe.
Often, when doing image splitting, it is necessary to know not only where the main subject is, but also to know the detail surrounding the main subject. For example, the artist may define and store a recipe requiring that a template with three window openings have a main subject face in a larger left window and low color non-face and non-main subject in the two smaller windows to the right. Image segmentation algorithms provided on the computer system along with people or main subject detection algorithms are oftentimes adequate to fulfill aesthetic image splitting requirements depending upon the artist's template recipe. A more advanced splitting technique uses seam carving as described by A. Shai, S. Ariel in “Seam Carving for Content-Aware Image Resizing”, ACM Transactions on Graphics, Vol. 26, issue 3, July 2007, to determine low energy regions of the image. This document is incorporated by reference herein in its entirety. These low energy regions are ideal candidates for splitting and can significantly aid the creative designer to define template recipes for attractive coordinated image displays.
Stored template recipe files created in advance by an artist can apply to the overall image enhanceable item 300, as well as for each window opening within the item. Each window recipe file can be combined and stored with other window recipe files to generate a stored template recipe file. These recipes can be designed and executed on the computer system for determining parameters such as, but not limited to, landscape vs. portrait orientation, colorfulness, sepia, color tint, person, face, main subject, size of person/face/main subject, location of person/face/main subject, surrounding regions of person/face/main subject.
In a preferred embodiment of the present invention, in order to size objects correctly, a relative measurement algorithm is used. This programmed system can specify size and location of an object in an image via a variable measure called a unit. A unit can be defined as the width or height of a face, the width or height of a main subject, or the width or height of an object in the image. A window recipe file can specify requirements to fill a window with a single person, that a face is 20% fill; where face location is specified as two unit spacings from a left edge, one unit spacing from a top edge; where there is low texture in the area of the image including two units of spacing to the right of the face or main subject; where the colorfulness of the person's shirt/blouse/dress (at 2 units below face) is high. The capability to specify and store in the computer system these required image attributes as window recipes for later repeatable use provides extensive creative flexibility for an artist.
In another preferred embodiment of the present invention, images can be scored on a merit system. The image that scores highest for a particular window recipe is indicated as assigned to the corresponding window. This can be programmed to occur automatically upon activating an application on the computer system, provided that there is a number of computer accessible images available, with no user interaction. This frees the user from the laborious and time consuming task of looking through many image files and is able to automatically create an image enhanceable item 300 that is often much superior to what a lay user could create because the designed look was created by a skilled artist. Of course, if the user is not happy with the automatic selection of templates and images, he/she can manually override the system, by selecting a user control option provided by the program, and either rescale, replace, re-crop, re-enhance, or modify an image using various off-the-shelf software products for digital image editing, or designate a different image from his or her collection of images for a particular virtual template window opening.
An exemplary recipe for a window, which can be one of a plurality of windows in an image template, is illustrated by the following code segment:
In the above recipe, each required image attribute entry is delineated by brackets and, where appropriate, is followed by an entry defining its weighting factor. In the first line, Orientation is the image required attribute, and the value paired up with Orientation key is the value 1, which in this example represents landscape orientation (“0” indicates a portrait orientation). Thus, landscape images will receive an attribute score of 1, portrait images a score of 0. For this example and the others that follow below, the ordering of the entries is arbitrary. The weighting factor is used for scoring images as is explained below.
The FacialCloseToNum required image attribute indicates that the input image should contain close to three faces according to its value. An image with exactly three faces would receive a score of 1. Images with 2 or 4 images receive a score of 0.5, images with 1 or 5 images receive a score of 0.25, images with zero faces receive a score of 0, images with 6 images receive a score of 0.125, images with 7 or more images receive a score that can be calculated by score=1/(2^(ActualFaces−DesiredFaces)), where DesiredFaces is the Datum value of 3 in the FacialCloseToNum attribute line. Similar attribute tags can be programmed, using any range of assigned values or by applying various value formulas, for FacialNumExact (an exact number of faces in an image), FacialGreaterThanEqualToNum (for a minimum number of faces in an image), and FacialLessThanEqualToNum (for a maximum number of faces in an image).
The Favorite tag defines whether the required image attribute indicates using one or more of a user's designated favorite image or images. Favorite images will get a value of 1, non-favorite images will get a value of 0. Of course, other value assignment strategies can be programmed into any window or template recipe.
The resolution tag defines a resolution based on pixels required for the image in the corresponding window. In this case, the image is required to have an image attribute resolution of 2431800 pixels (computed as (width in pixels)×(height in pixels)) or higher to obtain the resolution weight added to its overall score. Images with total number of pixels>=the Datum attribute receive a value of 1, all other images receive a value of 0.
The weighting factor as used for required image attributes enables the artist to specify a different level of importance for each required image attribute when determining the image's overall score. The weighting factor multiplied by the image attribute score generates a weighted score for each attribute. The sum of all weighted scores gives a total score for each image evaluated by this computation method. Clearly, various procedures for determining compatibility of images with required attributes can be devised and easily programmed for use with this embodiment of the invention. The image with the highest total score is the image that most satisfies, or is most compatible, with the window recipe. In a preferred embodiment of the present invention, if a weight line is omitted for a particular attribute, the default weight of 10 is automatically assigned. Attributes which are more important have weights greater than 10, and attributes which are less important have weights less than 10. If an attribute is mandatory, a weight of 100 (or higher) is assigned. As used herein to modify the term “attribute(s)”, the terms “required” and “requirements”, etc., are variable requirements according to the weights ascribed to any particular attribute. Thus, an attribute rises to the level of “mandatory” if its weight is programmed with a value of 100. For example, in the above code segment, the ResolutionWeight is equal to 100. This means that only images with 2431800 pixels or higher can be used in this window opening.
All entries, except the last one in the above example recipe segment, are used for scoring candidate images. The last entry, in this example it's “OffsetLeft”, is used as a template/window specification, and not as an image selection guideline, and thus the last line does not have a paired weight line. The last entry specifies that the highest scoring image, or the image selected for a virtual template opening, should be modified in some respect. This can be referred to as a “post processing” step because it is used as a fine tuning step or as a post-selection layout step. In this example the specification states that the main subject be shifted left using the rule of thirds before virtual placement into the template opening. The rule of thirds is a photographic composition rule that has proven to yield well balanced or natural looking prints by the average observer, and can be programmed for use by computer system 26. If an image is broken into an equally spaced 3×3 grid with two horizontal and two vertical lines, the aesthetic design idea is to move or place the image such that the subject of interest is located on one of the four dividing lines. Similar rule of thirds tag entries include “OffsetRight”, “OffsetTop”, “OffsetBottom”, “OffsetUL”, “OffsetUR”, “OffsetLL”, and “OffsetLR” to place (or move) the image such that the main subject is located on one of the four rule of thirds lines or, as in the last four examples above, on one of the four intersecting points of these lines. Other template specification tag entries, for post-selection layout, exist to modify a specific main subject/face size and location, or overall image rendition, such as “sepia”, “B/W”, “colorful”, “light”, “high Contrast”, “soft”, “cartoonization”, “emboss”, and other artistic or aesthetic effects.
In another preferred embodiment of the present invention, favorite or primary images can be designated as such in an images recorded metadata and used as either the focal point of the image enhanceable product, can be tagged for special or unique treatment either at pre- or pos-selection, or can be used as a primary template image such that images in other template windows must conform, as defined in their associated required image attributes, in some aspect to key characteristics of one or more image attributes of favorite or primary images. For example, a large window opening in the center of a template might be appropriate to place one of the customer's primary images. Such an opening can be designated as a primary opening in a template and its required image attributes can be defined such that it will receive an image designated as primary or favorite. A user of the computer system can be programmably prompted to identify such images during operation of computer system 26 or such images might have a “primary” indication already stored in association with the image as a recorded metadata tag. Other template openings can also be designated as primary, or as secondary, etc., and image files can include metadata information indicating which ones are favorite or primary images so that the programmed computer system can search and find these images for placement into the designated template window. As mentioned earlier, a customer can designate an image as a primary or favorite image, can dictate the location of their favorite image within one of a number of template windows, or may tag many images as favorites, letting the program automatically pick from a selection of favorite images using other image attributes as selection criteria based on required image attributes defined in template or window recipes. Other images in neighboring template openings can then be automatically selected by computer programming according to window recipes stored in association therewith. Other features of window recipes can include detected similar color and tone characteristics, or other stored image aspects, of the particular favorite image or images.
For the example recipe defined by the program segment above, an image would be required to have the following image attributes: it would have to be designated as a favorite image, be laid out in landscape mode, be a group shot, optimally with 3 people in it, and needs to have a resolution of 2431800 pixels or higher to obtain the perfect total score of 40 in this example (the resolution weight of “100” is not an actual weight because the number “100” triggers a special case handling in the program that the resolution weight is mandatory, and not just a relative level of a requirement). The total score is obtained by evaluating each attribute independently, then summing together each image's weighted values. Also in this recipe example, as dictated by the attribute weights, aspect ratio, with a weight of 20 is more important to the artist than most of the other scoring attributes which each have a weight of 10.
The above required image attribute values and weights (in XML these are referred to as key value pairs) are designed by the artist and are stored in computer storage or memory 40 as definition files for each template opening (i.e. “window recipe”) as a method to realize the artist's aesthetic intent. In addition to defining attributes on a window by window basis, openings can be referenced by a group name, or template name, where all window openings in a named group, or template, have at least one common attribute (i.e. “template recipe”). In addition, a grouping of all openings in a template can be programmed for enabling global required image attributes, where such attributes are required for all openings in the template. A precedence procedure can also be programmably defined, for example: required image attributes for a group of openings supersede global required image attributes, and individual opening required image attributes supersede group attributes. For example, the artist may define a global required image attribute requiring that all openings must contain people by specifying under the global recipe:
By weighting such an attribute with a value of 100, it becomes mandatory. Therefore, if there are no satisfactory images found after performing an automatic programmed search through available image data bases the program may notify the user that it can't be fulfilled. In the above global attribute example, if a weight were not explicitly specified, a default weight of 10 is automatically assigned in a preferred embodiment. Subsequent examples herein will not specify weight values for purposes of simplicity of description.
In another preferred embodiment, a window recipe might define required image attributes such that each window includes an image having a particular time-date stamp, GPS location, camera parameters, or other user definable tags. For example, a window recipe can include any combination of the following example required image attributes:
The first two examples above define the minimum and maximum date and time of capture of the image. DateMax and TimeMin similarly exist. In the above example for DateMin, the full month/day/year is specified. This format can be changed to any of a number of accepted formats like day/month/year. It is also acceptable to include partial dates. If the day of the month is to be omitted, the above Datum would be “10//2008”. The third example shows an example of specifying an exact date. A similar field exists for Time. In addition to specifying an exact date as above, well known holidays can be substituted, such as Datum=“Thanksgiving”, or Datum=“Memorial Day”. Similar substitutions of “Noon” and “Midnight” can be used for the time entry. The fourth and fifth examples require the image to be captured at a particular GPS location, such as a campground near a particular tourist attraction (within 75.5 meters in this example) in Yellowstone National Park; the fifth example allows control over the camera's image capture conditions such as ISO, and can be used to specify more advanced features such as focus distance, F#, etc. The final example can be used to require images with particular customer or third party provided metadata tags. In this example, recorded metadata associated with an image is required to include the text “Disney”. For another example, it is possible for a user to specify for a window recipe only shots of “Uncle Joe”, or for a Kiosk application to require looking for images on a particular website that are metadata tagged for specific usage on a particular template or a particular family of templates. For example, the program can be designed to look for images tagged with “Uncle Joe” on the user's or users friend's Facebook account, and such an account on a computer system accessible web server can automatically be searched and images having the required metadata retrieved; or images from Flickr that are tagged as “sunset” can be required in a window recipe; or custom databases can be created locally on the computer system or on a coupled computer system in a distributed network, or on network accessible servers with the intention of combining images stored in them into user enhanceable products.
Public or private image databases may have tagged images with specific content available for usage in the user enhanceable product 300. This content may be generic such as “beach”, “palm trees”, or may have a unique identifier with a product label such as “Kodak Premium”, or may be authorized by and associated with a particular celebrity such as a “Famous Designer”, or may be part of a special store promotion such as “Walmart Summer 2008”. More than one Title recipe entry can exist to tag a window opening, for example “Famous Designer Fall 2008 Collection” or “Colorful Leaves” is required to appear together with “Disney”. The usage of such template recipes can be used to entice consumers to purchase the latest celebrity branded and celebrity approved image enhanceable item 300 and may include the automatic diversion of royalties to said entity and entice the return of customers time and time again to purchase the latest offering by a specific celebrity/artist/sports/famous person or event such as a World Series 2009 collection.
Once again, the programming of global recipes and local recipes for each window can be used to generate recipes for templates of various types. For example, when creating a picture book, it is often desirable, but not mandatory, to arrange pictures in chronological order. We can create global recipe for a picture book template to “softly” enforce such rules for each window in the virtual picture book template, as shown below, by assigning a lower weight values for “soft” enforcement:
In the above example, the recipe requires that W1 (window 1) should have a date (or time stamp) that is greater than or equal to W0 (window 0), a greater time stamp meaning later in time, where W1 and W0 are more generally described as W<integer>, where integer is any window opening ID. By modifying the assigned weight of the above recipe, we can enforce more or less the desire to maintain a chronological ordering of customer images in a picture book as compared to other image attributes. To force chronological ordering as a mandatory entry, the weights would be set to 100 for this entry for each window opening.
In another preferred embodiment of the present invention, the window recipe might include color intent such that each window opening in a template has a defined stored color intent. For example, a recipe can control color content as follows:
The first entry defines the required color attribute via values such as “warm” and “cool’ for the image; the second entry defines the allowable color attributes via specific color values as “red”, “green”, “blue”; the third example defines the allowable color attributes via specific 24-bit RGB code values to define required image attributes with a specific color value to match, for example, a company logo, pantone color, or other pre-defined color scheme. The final line is the most complex: if the attribute value (Datum) is a W<integer>, this indicates that the primary color from an image disposed in window opening <integer> is first to be extracted, and then the current window opening is defined to require image attributes to have primary colors similar to, or dependent upon, that of the image in window <integer>.
In the previous examples, images that have attributes as defined in the required image attributes score higher than those that do not. After the best image is selected, the image is further modified to accommodate the window recipe command. So, for the last example, if W0 had a warm primary color, the image that matches this warm color the closest gets the highest weight. Upon final selection of an image, the image selected for this window opening is further checked and modified, i.e. colors are modified to warm properties, if necessary, to obtain this same warm primary color from window W0.
In another preferred embodiment embodiment of the present invention, the window recipe might include tonescale intent such that images in each window opening in a template has a predetermined tonescale. For example, each recipe entry line can control tonescale in any of the following ways:
The first example controls the allowable image scene contrast attributes via values such as “light” and “dark”; the second example controls the allowable image contrast intent via specific attribute values such as ‘high, “low”, or “normal”. The next example is similar to the color intent. If the required image attribute value is defined as W<integer>, this indicates that the primary tonescale characteristics from an image in window opening <integer> is first to be extracted, and then the required image attributes for an image in a current window is chosen to have a tonescale similar to, or dependent upon, that of an image in window <integer>. This feature can be used in conjunction with or independently from the designation of a favorite image, where the color/tone information is extracted from the favorite image and used in other window openings to define required image attributes. This is a powerful program extension, for example, that would allow, for example, the color and tone of all images in a template to match a certain reference image. The certain reference image might or might not be included in the template. The reference image might be of a paint swatch, fabric, or an image of a room or any other item.
This recipe type can also be used in other commercial applications. For example, a window recipe can be created for a school yearbook whereby each graduate picture is a window having defined required image attributes associated therewith, and the color and tone of each picture be made to match, or be compatible with, a reference image. Similar applications exist for advertising. A mail order catalog or weekly store circular can use window recipes to describe page layouts such that all or a group of images have similar contrast. A programmable storefront display can be advantaged with similar required image attribute designs.
In the third example, if the required attribute value is ˜W<integer>, this indicates that the primary color/tone characteristics from an image in window opening <integer> is first to be extracted, and then the current window opening is chosen to include am image having a color/tone opposite to that of the image in window <integer>. So, if the window <integer> image is high contrast, the current window opening will have an image with low contrast. Accordingly, the higher the contrast in window <integer> image, the lower the contrast in the current window image.
If the attribute value is (W<integer1>|W<integer2>| . . . |W<integern>), this indicates that the primary color/tone characteristics from images in windows <integer1> . . . <integern> are first to be extracted, and then the required image characteristics in the current window opening is chosen to have a color/tone similar to the average of the images in the n window openings. Similarly, if attribute tag value is ˜(W<integer1>|W<integer2>| . . . |W<integern>), this indicates that the primary color/tone characteristics from images in window openings <integer1> . . . <integern> are first to be extracted, and then the current window opening is chosen to have an image with color/tone opposite to the average of the n window openings.
Using the color and tonescale required image attribute values, the artist can design all window openings to have similar color/tone content, unique color/tone content, or some combination thereof. It should be readily understandable that these two examples can be modified to include other required image attributes such as texture, colorfulness, season, time of year, time of day, people, sharpness, and noise.
In situations where a customer manually adjusts the color/tone/sharpness/cropping of an image in a window opening via the user control option that affects the color/tone/sharpness/cropping of another window in the project whose required image attributes are dependent upon such characteristics of the image, the dependent image can be programmed to be automatically adjusted accordingly. For example, if window opening W5 is defined to have an image with identical contrast as the image in W1, and the user then manually adjusts the contrast of the image in W1, making it higher, the contrast of the image in window opening W5 will also increase automatically according to the program. In some situations, the user may choose not to activate this feature, so the user control option can be configured accordingly.
In another preferred embodiment of the present invention, the window recipe might include further object detection or scene reflection required attributes. For example, the window recipe can dictate more specific requirements for each window opening:
The first example controls whether the window opening should have a specific object in the scene (a balloon), and can also specify other objects such as “BirthdayCake”, “SoccerBall”, or “Dog”. The second example controls a required attribute defining a theme for the opening with a value such as “city”, “colorful”, or “dreamy”. The former example can be used to search for and select all images from a soccer game or a birthday party. The latter example can be used to define a proper mood for a particular window opening. The last line utilizes a similarity measure between images. It defines the image in the current opening to be similar to that of Window opening 1. This feature supports the attribute which indicates “dissimilarity” in images, and the ‘|’ feature which indicates averaging several windows together. The similarity measure can use any number of image attributes, but RGB histogram statistics such as Chi-squared analysis or low resolution pixel correlation metrics can also be used, are fast, and perform quite well. Two low resolution metrics that perform well are Mahalanobis distance and edge map correlation. The Mahalanobis distance is similar to standard deviation, but also takes into consideration the correlation between each of the individual color channels. The edge map correlation is performed on low resolution images (64×96 pixels or smaller) by comparing the derivative (using Sobel filter) of the reference image to the derivative of all other images. More complex similarity metrics which utilize complex semantic information can be readily incorporated.
In another preferred embodiment of the present invention, the recipe can include further object, and scene attributes, such as in the following examples:
The top example controls the ability to specify a generic location required attribute value such as “water”, “beach”, “city”, or “indoors”. These locations are overall scene classifications, and the resulting classification is based on an output from a scene classifying computation carried out on a digital image. The last example can further constrain the image selection procedure by requiring that scenes in an image contain certain objects or characteristics such as “snow”, “water”, “sky”, “foliage”, “dirt”, and “skin”. The latter example can be implemented via belief maps. If no snow is found in an image, a value of 0 is returned as that image's attribute value for “snow”. The more snow found (the greater the percentage of snow pixels), the higher is its corresponding value for the required “snow” attribute in the window recipe. A discussion of sky detection is provided in Gallagher, et al., “Detection of Sky in Digital Color Images”, U.S. Pat. No. 7,336,819 and an example of foliage detection is provided in “Color Segmentation as an Aid to White Balancing for Digital Still Cameras”, by Cooper, Proceedings of SPIE, v4300, p 164-171, 2001. Both of these references are incorporated herein by reference in their entirety.
In another preferred embodiment of the present invention, the recipe can include further people attributes, such as in the following examples:
The first example attribute requires that the age of people depicted in the image be less than or equal to a certain number, in this example less than or equal to two years old. Similar attribute requirements exist for AgeGreaterThanEqualtoNum, and AgeClosetoNum. Often we want people facing “forward”, but, it is often desirable for a person or persons pictured in one window to be looking in the direction of a second window opening. The second example requires a particular facial pose of people or a person depicted in the image, in this example the person should be looking left. The next example above indicates a preference for a facial expression or demeanor. This allows control over facial expressions such as “angry”, “sad”, “surprised”, “happy”, “smile”, and “neutral”. The last example can be used to make sure the eyes of all found faces are open. Similar tags can be written to control related features such as ethniticity, gender, hair color/style, eye color, skin tone, height, age, facial or hair accessories (sunglasses, headband), facial hair, and mouth open/closed. These controls are well known algorithms familiar to persons of skill in the art, and are similar to those previously disclosed.
The next example utilizes facial recognition to identify a person in an image.
In the first example, the required personID attribute controls which person's image is being inserted into an opening. The attribute value “Uncle Joe” can be satisfied either by searching for and finding recorded metadata associated with the image, or by utilizing an accompanying facial recognition database and extracting the metadata. It is well known that extracted metadata can thereafter be recorded with the image and need not be obtained by repeating the extraction processing later. Many cameras and imaging software packages now include facial recognition capabilities. The last example above utilizes the window relationship feature for window recipes. This last example says that the image being placed into the current opening must have a person, and that person must also be in window W0. This last example is a generic way of describing a relationship of people appearing in more than one template window. For example, we can have a center template window contain a closeup of a single person as dictated by a favorite image being placed therein, and all surrounding template windows may thereby be defined to require that same person somewhere in the image.
In another preferred embodiment of the present invention, the recipe can include further image selection steps, such as in the following examples:
The first example attribute defines the window opening width over height aspect ratio. In this case, we know the width is 1.5 times the height. When used together with the second example, defining an oval, we can define a maskable area to overlay on top of any candidate images. We then use the maskable area to score each of the candidate images. WindowShapes can include Rectangle, Circles, Ovals, Pentagrams, Hexagons, RoundedCornerRectangles, and can be easily extended to any other shape. When AspectRatio and WindowShape are used in combination with other recipe items such as PersonID, FacesClose2Num, etc, only the maskable area is included in the overall score. For example, a 1.5 aspect ratio rectangle may include 5 faces, while a 1.5 aspect ratio oval may only include 3 faces and would result in different scores for the same underlying image.
In another preferred embodiment of the present invention, the recipe can include multiple image selections per window opening. This is defined as having two or more recipe entries for each window. In the case of overlap, the last recipe entry is the top entry unless the Order entry is used. For example, we may have the following two examples for a single window opening:
Along with the following second recipe for the same window opening:
The first recipe specifies an oval mask, half as wide as tall. We want exactly two people, Uncle Joe and Aunt Suzie. The second recipe also specifies an oval mask, half as wide as tall, but, it is an inverse mask. So, the image area outside of the oval is used for image scoring. In this case, the second image outer mask area image should have no people and contain sand.
The above two examples assume the image is centered and fills the window opening as much as possible. It is possible to do various collage type entries for each opening, by specifying the subwindow [X,Y] offset and the subwindow width and height for each of the recipes that fulfill a particular opening.
In the above example, the image that scores highest for this particular recipe will be fit to the lower half of the window opening. Values for SubWindowX, SubWindowY, SubWindowW, and SubWindowH scale between 0 to 1.0, where (0,0) is the upper left corner and (1,1) is the lower right corner, but other conventions can be used. Finally, these subwindows can be feathered into one another and order can be assigned as follows:
In the above example, feather command controls the feathered edges on an image. The convention used here is that the feather value can be any percentage of the window opening width. The Order value assigned a floating point hierarchy to each image in a window opening, where higher values move the window further to the front. If no Order value is specifically stated, or if we have two identical Order entries, it is assumed the order of the current recipe is in front of the previously found recipe for the same window opening. If the first recipe per window opening does not have an Order value, a default of 0 is used, which means it will be in the back, and all other images will be placed in front of it. Using masks, inverse, masks, locations, and order enable a powerful scripting language for the skilled artist.
In another preferred embodiment of the present invention, the recipe can include the demarcation of specific regions of the image. In particular, region segmentation of sky, snow, skin, foliage is often determined by color and location within the image. Background areas are often denoted by lack of texture. These regions can automatically be masked off and combined or replaced with other attribute areas:
The first example can be used to set the transparency of the non-sky region of the selected image to 0%, or set the transparency of the sky to 100%. The Datum value indicates the transparency value. When this image is overlaid on top of a previously fulfilled recipe for the same window opening, the sky will be replaced with the image contents of the background image. When the Select tags are used, all areas not included are set to a transparency of 100, such that they will not be seen in the reproduced image. In the second example, we are selecting just the facial region of an image, the rest of the image will be transparent. The face will have a 50% transparency, such that it will be blended with the existing background image at 50%. When used with SelectFacelnverse of a second image, we can swap the face of one person with another. The results of the face swap are much better if the exact facial outline is selected using one of the previously mentioned examples. Finally, in the last example, the background of an image is selected based upon image energy, where energy is described by Avidan and Shamir in “Seam Carving for Content-Aware Image Resizing” in ACM, Vol. 26, No. 3, July 2007, incorporated by reference herein in its entirety. The low energy area is considered background area, and in this case, the color of the background area will be made to match the primary color of what was in window opening 0.
The first example can be used to create adaptive montages of either aesthetic or practical significance. Aesthetically, special effects can be created, for example, in an advertising poster to add an image or text in a sky or background area of an image. The second example can also be used to create artistic effects, but can also be used to replace faces that have inappropriate facial expressions, eye closers, head pose, etc, with faces that are more appropriate. The second example can also be used in commercial purposes to replace, for example all the heads of models in the store circular with the children in ones family. A custom softcopy circular can be created for each household, not only with custom targeted content, but with family members wearing the clothing, operating equipment, or looking on in the distance. Social websites like Facebook can use such recipes to create custom slideshows for both pleasure and profit. The latter example can be used in context with the previously described adaptive recipes such that the end product more closely follows a theme or more closely resembles the artists intent. For example, the background information can be made to have the texture or color of a favorite image, made transparent or just driven to white so that the window opening can concentrate on the main subject.
In another preferred embodiment of the present invention, the recipe can include the usage of units in window recipes to specify the size and location of a particular face, main subject, or other object. For example:
The first recipe line defines the current window required image attribute needs an image with one face. The second line says this required image attribute is mandatory—it is mandatory for the image to have exactly one face. Images containing 0 or greater than 1 face are not considered for inclusion in the current window opening. If no images with exactly one face are found, the recipe cannot be fulfilled, and (as with all other unfulfilled mandatory required image attributes) the job can be terminated, a notification can be provided to the user, an option to override the requirement can be provided, or an option for the user to insert a selected image can be provided, or the user can be prompted for some other type of intervention. The third line defines that this face must occupy 50% of the image area. Images with slightly smaller or larger faces will score high, images with large face differences will score lower. One formula used is 1−abs(TargetFaceSize−ActualFaceSize), but non-linear variants such as logarithmic, gamma, and similar functions can also be incorporated. Regardless of the preferred variations for evaluation, an image with one face, covering closest to 50% of the window area will be calculated to score the highest and be selected for inclusion in the template window. After selection, the image can be specified for post-processing and be resampled such that the face will occupy exactly 50% of the window area. The final line above specifies to center the face in the middle of the window opening. The FaceSize and FaceLocation key value pairs include the W0 and similar options such that the size and location of faces can mimic that of the favorite images placed in the template. In addition, because sizing is so critical, the FaceSize and FaceLocation key value pairs include the more advanced W0 notation, such as:
This defines, for a window recipe, that the face in the current window opening should be 80% of the size of the face in window opening W0. When multiple faces occur, the FaceSize and Location refer to the average face size and centroid of all faces. When W0|W1|W2 notation is used, it defines using an average size face from window openings W0, W1, and W2. In addition to “Center”, FaceLocation can include the rule of third locations, UpperLeft, UpperRight, LowerLeft, LowerRight, as well as Upper, Lower, Left, Right. The W0 notation for FaceLocation includes mirroring the location such as:
The above recipe line will mimic the horizontal location of the face—so if the face in W0 was in the upper third, the face in the current window would occur in the lower third of the image. To make relative locations as compared to W0, offsets are used such as:
Where the above line says the face location in the current window should be the same as the face in window W0, but, should be offset 20% to the right and 30% towards the bottom. The above commands are illustrated for use in detected faces, but equivalent commands exist for main subject size and location as well as object size and location when the main subject is something besides a face.
The process of evaluating the fitness, or compatibility, of multiple images as against a window recipe for a particular window is shown in
The first listed attribute is for favorite images. 620 and 640 receive a value of 1. 610 and 630 receive a value of 0. Multiplying by the favorite weight of 20, we get weighted values of 0, 20, 0, 20, respectively, for each of the four images shown.
The FacialCloseToNum attribute favors images with 2 faces. Unfortunately, none of the images have two faces—each image having one too many, or one too few faces, giving a value of 0.5 for each image. Since the FacialCloseToNum weight was 40, multiplying 40 and 0.5 results in a weighted value of 20 for all four images.
The FacialPose attribute favors all frontal poses. Facial pose estimators return the pitch and yaw of a face in degrees. The value is calculated by setting value=(100−(Pitch+Yaw))/100, and then clipping such that the value calculation is normalized from 0 to 1.0. This strongly favors frontal poses where pitch=yaw=0, and penalizes poses otherwise.
The Location attribute is looking for indoor scenes. State of the art indoor/outdoor locators use metadata tags of ISO, exposure time, and flash fire, along with key semantic information such as foliage and sky colored pixels, along with overall color temperature of the scene. The returned value is an estimate that the scene is indeed an outdoor scene. For scenes 1, 2, 3, n, the value returned was 1, 0.7, 0.25, and 1, yielding weighted values of 9, 7, 2.5, and 10.
With the evaluation of the above weighted values for each image and attribute completed, we sum the weighted values for each image, giving a total of 49, 67, 42.5, and 55 for images 1, 2, 3, and n and assign to 615, 625, 635, and 645 in
The last two lines of the recipe in
Additional post-processing instructions can be utilized to change the color, tone, sharpness of images and change the relative color, tone, contrast to other window openings as follows:
Where the first example defines applying an autofix or automatic image enhancement to the image in the current window. Similar entries exist for AutoNeutralBalance, AutoColorBalance, AutoContrast, AutoSharpness, and AutoZoomCrop. If AutoZoomCrop is used, and the window opening still scores highly if small, outlier faces are automatically cropped out of the image, they will be cropped out as long as the resulting image resolution is sufficient for the final enhanceable product. The second example indicates that the contrast of the image in the current window should be less than the contrast than the image in window 0. Similar entries exist for Colorfulness, Brightness, and Sharpness. In addition, the Datum values of “>W0” would mean the criteria in question would be greater than window 0; and “>>W0” would indicate the criteria in question would be much greater than window 0.
The concept of using the > and < characters in the recipe also allow relative specification of window content. For example:
Where the first example defines having the current window contain a person whose age is greater than the estimated age of the person in window 0. If >> was used, the age would have to be much greater. These relative relations can be easily defined within the program as desired, for example, “>” could be defined as five years, and “>>” could be defined as fifteen or twenty years. The next two examples show how to specify the age of the person in the current window to be between 10 and 30 years the estimated age of the person in window 0. The addition of >, >>, <, and << along with the offsets, |(averaging), and ˜(inverting) enable a rich suite of tools for the artist.
The described embodiments of the present invention are not limited to the values and attributes defined herein, and can be extended to other parameters as an artist sees fit. As computer and hand held device algorithms evolve, future available image metadata can be incorporated. While some metadata tags are easy to evaluate, such as time stamp, others are more complex, such as selecting images with balloons. As the recipe is loaded into the program, a list of necessary attributes is accumulated, and only those attributes that are required for a particular image enhanceable item 300 are resolved for each image.
Another preferred embodiment of the present invention is the usage of adaptive recipes. For example, a user is making over a room in his home and would like to hang decorative image enhanced items on the wall. The user can bring a paint sample from the room to a retail kiosk that is running a computer system program embodiment of the present invention such as on computer system 26, and the paint sample can be scanned by the kiosk's flatbed scanner 38. This paint sample can be used to define overall color and texture required image attributes for the decorative image enhanceable wall hanging. Similarly, a frame or matte can be scanned for defining attributes such as color and texture. If a flatbed scanner is not available to kiosk computer system 26, a calibrated webcam or video camera at a kiosk will suffice. Similar techniques can be used if a user brings in a swatch of material, such as a pillow, cushion, curtain, or any other object in the room to be decorated. Not only can the resultant user enhanceable product be made to match this material, the background matte of the template can be a replication of the actual material or paint sample. Finally, a user can take photos of objects or walls, and those objects can be used as the reference digital image. For example, a user may provide four reference materials, R0 can be a paint swatch, R1 can be a curtain, R2 can be a pillow, and R3 can be a sample from a magazine. The following window recipe examples can be applied to these samples.
In the first example, the color of an image in the current window can be made to match the color of R0, which is the paint swatch. This is done by first digitally balancing the image to neutral and then adding a color cast to the image—the color cast being a match to the color cast of the paint swatch. In the second example, the tonescale of the image can be similarly manipulated to match an average contrast of the curtain and pillow. In the third example, the matte color will be adjusted to match that of the paint sample. In the final example, the matte texture will be adjusted to match the texture sample from the magazine clipping. The above use of reference images is very similar to the window matching entries described above, say Datum=“W0”, but the key difference is that the reference image in these examples is not displayed in the completed coordinated image template design or in the final image enhanced product.
Another preferred embodiment of the present invention involves the evaluation of multiple window recipes on a single set of images. For example, a Famous Designer may come out with the Fall 2008 line of template recipes. Depending on the images provided by a user, one particular template recipe might be more appropriate than another. For example, if a customer loads a group of images containing sports scenes, a Famous Designer sport theme template recipe may be used initially to begin designing the final image enhanceable product. If the customer loads images of children, a Famous Designer infant or toddler theme may be selected. If a customer loads a group of images containing water scenes, a Famous Designer beach template may be chosen. The names of the template themes can also be displayed to a user so that the user can select and open a template initially whose name reflects the type of context that fits the user's image collection. The selection of a favorite image will also cause some templates to score higher than others. Based on a different selection of a favorite image, the templates that score highest can change drastically. The highest scoring template as calculated by the program can be used, or the top n highest scoring templates can be programmed to be presented to the customer for him/her to make a final decision as to which is his/her favorite template. The procedure that is followed to identify the types of scenes depicted in groups of customer images is well known in the art and is not described further.
If several favorite images are selected by the user, the corresponding highest scoring templates can be presented to the customer, or the top n templates for each favorite image can be presented, or the top n templates for only the top scoring favorite image can be presented. The types of such template recipes is virtually unlimited. Image enhanceable items 300, can include framed templates, clothing items, greeting cards, photo books, and other tangible goods; but also to softcopy goods as well such as cell phone, PDA, and computer screen savers, digital greeting cards, as well as digital picture frames, and automatic digital presentations used in store fronts, office lobbies, stadiums, theatrical shows, and for personal enjoyment.
In another preferred embodiment of the present invention, a unique template recipe is assembled on the fly and pre-selected recipes are displayed in a pre-determined or random order. For example, a continually running slide show on a digital picture frame (which can be a picture frame at home or a large electronic storefront window) may automatically extract photos from a public or private website, or from a local computer. Each and every displayed frame can be unique and can be programmed never to be shown again, if so desired. A preferred embodiment of the present invention includes a procedure of writing a continually looping program that modifies the recipe in each iteration. During each iteration, the top x compositions are displayed in a random or a predetermined order. When selecting images from online galleries such as Facebook or Flicker, or Photobucket, the number of high scoring template configurations can be staggering- and a limit can be imposed for each iteration of the looping program. The looping program can have nested loops, each changing one or more recipe entries in random or predetermined order. Fulfilled recipes can be stored in a database to prevent displaying the same recipe twice, or to purposely replay recipes in a preprogrammed or random order.
Template recipe entries can be configured by, and grouped according to, product line and can be upgradeable over time. A similar architecture can be used in many product lines, with a baseline product only offering a subset of required image attributes to be selected, a moderate product offering an extended set of selectable required image attributes, and a premium product offering a full suite of required image attributes. In addition, the attribute readers can be designed to be forward compatible by ignoring any attribute values it does not understand or can not evaluate. This can occur if, for example, a new artist's collection of templates uses attributes not recognized by a previous version of the program. For example, if balloon object extraction is not supported, it is simply ignored.
Potential designs for interpreting and using the corresponding attributes is numerous. At the highest level, a parameterized and weighted fitting algorithm scores each individual image from a set of required image attributes for each window in the template. The objective of the parameterized weighted fitting algorithm is to find the best (most appropriate) image for each window based on image/template image window required attribute comparison. As used throughout herein, a “required” opening attribute (or window attribute) is a soft or hard (depending on weights) requirement, defined in a recipe, that an image must satisfy. Image attributes are the metadata describing image characteristics that are extracted by digitally examining an image, or provided by a user, or generated by a camera, or otherwise provided and stored with an image or generated on the fly. The parameterized weighted fitting algorithm iterates through a list of windows and scores each image relative to required image attributes for a window and finds the best image for that template window.
The algorithm by which a preferred embodiment of the present invention is implemented is described in reference to
At step 810 the template was selected by the user and now the image selection begins. If the user has not yet provided images, either in a local computer accessible database, a network accessible database or on a portable or other storage device, then at step 820 the system will prompt the use to identify where images can be accessed. The prompt may or may not occur if a data base has previously been identified to the system or if a user has coupled a portable or other image storage device to the system. If the user indicates at step 805 that he will select favorite or primary images, then at step 830 he will indicate the selected images to the system and at this point the system will automatically proceed to step 905. If the user indicates that he will not provide favorite or primary images at step 805, then the system will automatically select images based on the step as defined in block 1210. At step 840 these automatically selected images will be digitally placed into the previously selected template. At step 815 the template and selected images will be displayed to the user on a screen or monitor and the user will have the option to accept the image product, however, the present invention includes an option 860 for the user not to review the image product. If the user accepts or if the user does not review the product, the product is output at step 870. If the user does not accept the output product at step 815, then the system repeats an automatic image selection using step 1210, but outputs a template plus images that is different from any previous outputs. Alternatively, a user control option 850 can be executed at this point wherein the user performs his own selection steps instead of the step 1210 performed automatically by the computer, and the user's selection is then output at 870. The user control option can be implemented via activation of an image editing software that is part of the image enhanceable product software or it can be implemented by a separate image editing software. The output step 870 operates according to the type of image enhanceable product that the user has selected. For example, and not by way of limitation, if a frame has been selected, then the output can involve a frame with selected images being delivered to the user (home delivery or retail outlet pick-up for example), or a page with the selected images may be printed which the user can insert into an actual frame that matches the digital frame depicted on the display.
At step 905, after the user has identified favorite or primary images on the computer system, the algorithm determines if the number of primary images identified is greater then the number of openings in the template. If so, then at step 1240 a subset, equal to the number of openings, of the favorite images whose image attributes satisfy required template attributes and template opening attributes, if any, are automatically selected and then in step 910, the template, together with the subset favorite images disposed in the openings is composed, and is displayed to the user for user approval at step 915. If an option to bypass user approval 912 is active then the template and automatically selected subset of the favorite images is output at step 916. Otherwise, if the user approves the template and images, the user indicates acceptance to the computer system of the template and images as displayed and the output step 916 commences. If the user approval step results in a non-acceptance by the user then the system proceeds back to step 910 to repeat automatic image selection as described above, except that the output of the automatic image selection step will be selected to be different than any previous output. As usual, programmed preference is given to higher scoring images that best match required template opening attributes, but if the user does not accept these, then lower scoring images will be used. If user control option 914 is activated then the user has the option to specify favorite images without reactivating the automatic selection step 1240. This can be performed by presenting to the user the template with current selected images displayed therein, and an option to replace the current images with other user selected images from an image data base. The user's favorite image selections are then used in the output step 916. At this point it is an optional programming step to limit the user option to only those images that are favorites, selected at step 830, or to present the whole user's image data base for selection.
If the number of favorite images is determined not to be greater than the number or template openings at step 905, then at step 925 the computer system determines, according to the programmed algorithm, if the number of favorite images indicated by the user is equal to the number of template openings at step 925. If so, in step 1230, each favorite image is assigned the optimal window such that the overall template score is maximized, then the images are placed in said window openings in step 940, and then output step at step 946. If an optional user control step 942 is active, the template and images disposed therein are displayed to the user to enable the user to select different templates or images, or to rearrange the images displayed. After the user has completed his selections, the output step 946 is activated.
If, at step 925, the computer system determines that the number of favorite images is not equal to the number of template openings, the algorithm selects additional images to compliment the favorite images using step 1220 which is described above. At step 930 these automatically selected images will be digitally placed into the previously selected digital template. At step 935 the template and selected images will be displayed to the user on a screen or monitor and the user will have the option to accept the image product at step 935. This embodiment of the present invention includes an option 934 which, when activated, allows the user to bypass review of the image product and go directly to output step 936. If the user accepts the displayed image product, or if the review bypass occurs, the image enhanced product is output at step 936. If the user does not accept the output product as displayed at step 935, then the system repeats an automatic image selection using step 1220, but outputs a template plus images that is different from any previous outputs. Alternatively, a user control option 932 can be executed at this point wherein the user performs his own selection step instead of the step 1220 performed automatically by the computer, and the user's selection is then output at 936. The output step 936, as do each of the other output steps described above, operates according to the type of image enhanceable product that the user has selected. For example, and not by way of limitation, if a frame has been selected, then the output can involve a frame with selected images being delivered to the user (home delivery or retail outlet pick-up for example), or a page with the selected images formatted to match the template openings may be printed which the user can insert into an actual frame that matches the digital frame depicted on the display. As described above, various image enhanceable products may be produced using the inventive methods described herein and the provision of these products can be completed at the output step in a variety of ways.
Referring to
If manual mode is enabled, the user may decide, at step 1005, to manually select one or more favorite images at step 1030 which are then passed to step 1040. If manual mode is disabled, the users images are automatically selected 1040. The highest scoring template is chosen 1040 by pairing the image set with the available templates in the system. In this instance, the score is not associated solely with the template, rather, it is a score generated by compatibility computation as between image attributes and the required image attributes defined by each of the window recipes in the template recipe. If a user has selected some of the template options, these are not modified by the algorithm. If a user has picked favorite images, these images, or a subset of these images, can be required by programming to be included in the final user enhanceable product and the highest scoring template is selected based on these favorite images being included in the template. Upon completion of template selection in step 1040, if automatic mode 1060 is enabled, the final template is selected, 1070. If automatic option is not turned on, the user has the ability to override 1050 the template selection process until he or she is satisfied. Once the user is satisfied, the final template is selected 1070 and the system now needs to fill this template with images and migrates to decision 1125 in
With regard to
Regarding the automatic template selection process as shown in
If a user decides at step 715 to choose their own template, the user can specify the exact template by scanning a template UPC barcode corresponding to a user enhanceable product available at a retailer that has installed a compatible kiosk, or itemized menu or the user can narrow their choices using an interactive interface at step 740 to help choose all the template options. If the user does not specifically provide the exact template or does not provide all template options 725, there is still some ambiguity as to which template that will be used. In such a case, the default template options 720 are superseded by the user defined options 760 and passed 780 on to the automatic template selection module starting with
If the user has provided all the template options at step 725 this results in a specific template to be used 750. We then enter the image selection phase starting at 810 in
The selection of favorite images 830 involves the user previewing, for example, thumbnail size images on the display screen and marking favorite images. The user can select a number of images for n template openings in the previously selected template, where n is equal to the number of window openings in the product. For each image that is split, n is decremented by 1 for each window that the split image occupies. To simplify the description of the algorithm in this embodiment of the invention, we will assume there is no image splitting. In addition to selecting their favorite images, the user is given the option of selecting their overall favorite or primary image. If a primary image is selected, it generally will be displayed in a prominent fashion on the user enhanceable product, and other template opening required image attributes can be based upon the primary image attributes (color, sharpness, person ID, scene content, etc.).
At step 905,
The procedure performed at 940 is as follows. The algorithm analyzes the images and determines what attributes they have either by accessing the metadata or semantic information extracted from the image or both. Both the metadata and semantic information can be calculated and stored with the image, or in a separate file. The fitness score of each image to each window opening is calculated and then the program initializes the template by assigning the highest scoring image to each window opening. If one image has the same score on two window openings, the priority is assigned to the lowest numbered window and the image is assigned to that opening. This presumes that the windows in the template are numbered. For example, window 0 (“W0”) is stored, by default, as the highest priority window, window 1 (“W1”) is next highest, and so forth. Windows for each template can be assigned numbers internally in a template file or in a window file. If two different images score identically high in one window opening, the image can also be assigned based on metadata fields that record the number of times an image was accessed, randomly, or by another selection method. An overall sum total score for the entire template mapping is computed and is the baseline starting point score before the algorithm iterates through the list of template openings for scoring other image combinations in the windows to determine if the baseline is exceeded. If another combination of images in the window openings results in an overall score higher than the baseline score, then that combination becomes our new highest scoring combination and we update our baseline score accordingly. We proceed until another combination again exceeds the current score or until iterations are programmed to end.
The algorithm updates the overall sum total score after each iteration through the list of template openings and images and can run through all possible combinations, until a set time period expires, or until some other programmed stop. The program can iterate through all images, or a subset, comparing a candidate image's attributes with some or all of the template openings' required attributes and computing new candidate image scores. When a candidate image's new fitness score is higher than a currently assigned image's score the algorithm enters into a potential swap condition.
Once the potential swap condition occurs, the image window's existing image can be scored against the candidate image's currently assigned image window and if both images score higher for both image windows then the swap can be selected. Swapping only happens if both images each score higher in the other opening and increase the template's overall score.
The algorithm will, at some point, contain a list of image windows H and a list of images I. A mapping of images to image windows where I1 maps to H1, In maps to Hn results in a fitness measurement score. Each image to image window mapping will be scored based on image attributes and required image window's attributes weighting factor. The algorithm can be selected to run for a computed number of iterations (number of images x number of holes or a subset thereof), until a predetermined fitness score is achieved, until the improvement from iteration to iteration is below a predetermined threshold, or until some other programmed cut off.
For each image window opening, the following are programming options:
The algorithm can swap the existing image Ii with every image Ir in the list of images and perform a weighted attributes fitness measurement score calculation by adding up the attributes weights that the image satisfies and determine if the Ir is a better fit with image window Hi.
If, for example, Ir to Hi is a better fit then the program can compare both images' (Ii and Ir) scores with their new possible swapped image windows (Hi and Hr) to determine if the average weighted attributes fitness measurement score is overall better. If the average fitness measurement score is better the two images' window openings are swapped, if it not, they are not swapped.
After the processing has occurred for each image window opening in a list, a total score for all image to image window mappings will be computed by summing the entire list of image scores and then comparing with the previous total score. If the newer score is better than the previous solution's score, the previous can be replaced with the newer candidate solution.
If we define imageList as the list of input images and imageWindowList as the list of window openings, the following algorithm can be used to iteratively find the optimal image to window opening configuration when the number of images is the same as the number of window openings:
Upon completion of step 940, if automatic mode 944 is enabled, the final user enhanceable product can be assembled, delivered, printed, etc. 946. If automatic option is not turned on, the user has the ability to override 942 the choices made by the image selection process until he or she is satisfied. For example, the user may not like the automatic rearrangement of images done by the automatic algorithm, or it is possible the user may want to select a new image and insert it into the product Once the user is satisfied, the final user enhanceable product can be assembled, delivered, and/or printed 946 as explained above. Referring back to step 905 of
There are multiple ways to select the optimal image to window opening configuration. The simplest approach is to evaluate the score of each image to each window opening and then assign the highest scoring image to each window opening. If one image has the same score on two window openings, a priority is assigned to each window opening and the image is assigned to the highest priority window opening. If two different images score identically high in one window opening, the image can be assigned based on metadata fields that record the number of times an image was accessed, randomly, or by another programmable procedure.
More sophisticated approaches can be used to maximize the overall template score as a whole. For small m and n, it is reasonable to evaluate all combinations and select the template along with m images arranged having the overall highest fitness score. For large m or n, a linear regression mapping m images to n window openings, where m>n, such as linear least squares or more complex non-linear or iterative approaches can be used, as described below.
Once the top m images are selected, and if automatic mode 912 is enabled, the final user enhanceable product is output 916. If automatic option is not turned on, the user has the ability to override 914 the choices made by the image selection process until he or she is satisfied, at which point the final user enhanceable product 916 is output or assembled, as the case requires.
The third possibility in
The simplest approach is to evaluate each candidate image to each of the remaining m−n window openings and pick the highest scoring image for each window opening, resolving ties by utilizing a priority with each window opening. When the image selection to pick from is obtrusively large, collecting semantic information for each image can be quite time consuming. A preferred embodiment is to only use recorded metadata to evaluate the fitness of each image as this is fast and can be done on any number of images in a quick fashion. The top x %, or the top p, where x or p is a user defined parameter, images are then selected for further interrogation, by evaluating the semantic information along with the metadata when evaluating the fitness score of each image. Another embodiment of the present invention is to analyze images in a hierarchical fashion. In the first layer, all images are evaluated by metadata alone. The top x % or p images are passed to the second round. In the second round, simpler, or fast semantic algorithms, such as color or histogram semantics are evaluated in the second round, where the top y % or q images are passed to the third round. In the third round, slightly more complex semantic algorithms such as face detection are evaluated and the top z % or r images are passed to the fourth round. In the fourth, or final round, the most complex semantic algorithms such as facial recognition, custom object detection are performed. This can be extended to any number of rounds. As we go from the first to last round, the number of candidate images shrinks, but the computational complexity of each analysis increases.
Additional programming features can be optionally implemented as follows:
In step 770, all template selection items were completed by default. In step 780, some of those default selection items were overridden by the consumer. For example, for step 780, all template options may be the same as in step 770 except that the user may override the number of window openings by reducing the total to two. Thus, in
The highest scoring template is chosen 1040 by pairing the image set with the available templates in the system. If a user has picked favorite images, these images, or a subset of these images can be required by programming to be included in the final user enhanceable product. If a “hero” image is marked, the template selected must have a “hero” window opening that is highly compatible with the “hero” image selected by the user. When scoring each template, each template is initially set to zero. It is possible for a manufacturer or retailer, to bias the creation of some templates over others by purposely initializing such templates at a higher score. For example, templates that generate more royalties, or that are faster to process, or that yield higher quality products may be preset at higher levels to encourage their usage over less desirable templates. Similarly, templates that are not desired may be given negative scores. If a single template is always to be used, it is given a score of positive 1000. Of course, if two or more templates are prescored at 1000, the one with the highest score after image evaluation is chosen. In the case of a tie, all templates can be assigned a numeric priority, and the template with the highest priority is chosen.
The decision of choosing the highest scoring template can be iterative. Each template in the system is evaluated one at a time across all images. To speed execution, this process is hierarchical. If a hero image is selected it is first evaluated across all templates. The top x %, where x is an adjustable/selectable parameter, templates are evaluated further, and the bottom (100−x) % templates are discarded. After the hero selection, the favorite images are scored in each template in each window opening. The highest scoring configuration for each template is stored. If n>m (if user favorite image is greater than window openings), then the selection is complete. If not, the program picks the top y %, where y is an adjustable parameter, and templates are evaluated further, and the bottom (100−y) % templates are discarded. Once again, for the remaining window openings in each remaining template, each image is evaluated and the top scoring templates are recorded. This hierarchical approach greatly speeds up program execution, especially when there are many templates, and many images to evaluate.
After the top scoring template is chosen, in automatic mode 1060, this template is selected by the algorithm 1070 and it along with its chosen images are passed to step 1125 in
Referring now to
The content of image, video, and audio assets can be analyzed using face, object, speech, and text identification and algorithms. The number of faces and relative positions in a scene or sequence of scenes can reveal important details to provide a context for the images. For example a large number of faces aligned in rows and columns indicates a formal posed context applicable to family reunions, team sports, graduations, and the like. Additional information such as team uniforms with identified logos and text would indicate a “sporting event”, matching caps and gowns would indicate a “graduation”, and assorted clothing may indicate a “family reunion”, and a white gown, matching colored gowns, and men in formal attire would indicate a “Wedding Party”. These indications combined with additional extracted and derived metadata provide an accurate context that enables the system to select appropriate images, provide relevant themes for the selected images, and to provide relevant additional images to the original image collection.
In the embodiment of
| Number | Name | Date | Kind |
|---|---|---|---|
| 5459819 | Watkins et al. | Oct 1995 | A |
| 5815645 | Fredlund et al. | Sep 1998 | A |
| 5986671 | Fredlund et al. | Nov 1999 | A |
| 6285468 | Cok et al. | Sep 2001 | B1 |
| 6389181 | Shaffer et al. | May 2002 | B2 |
| 6476903 | Slater et al. | Nov 2002 | B1 |
| 6636648 | Loui et al. | Oct 2003 | B2 |
| 6711291 | Stubler et al. | Mar 2004 | B1 |
| 6748097 | Gindele et al. | Jun 2004 | B1 |
| 6813618 | Loui et al. | Nov 2004 | B1 |
| 6822756 | Cok et al. | Nov 2004 | B1 |
| 6868190 | Morton | Mar 2005 | B1 |
| 6956671 | Monty et al. | Oct 2005 | B2 |
| 6958821 | McIntyre | Oct 2005 | B1 |
| 6967667 | Dworsky et al. | Nov 2005 | B2 |
| 6992787 | Fredlund et al. | Jan 2006 | B2 |
| 7000192 | Allen et al. | Feb 2006 | B2 |
| 7003731 | Rhoads et al. | Feb 2006 | B1 |
| 7035467 | Nicponski | Apr 2006 | B2 |
| 7043053 | Patton et al. | May 2006 | B1 |
| 7092966 | McIntyre | Aug 2006 | B2 |
| 7340676 | Geigel et al. | Mar 2008 | B2 |
| 7362919 | Das et al. | Apr 2008 | B2 |
| 7533129 | McIntyre | May 2009 | B2 |
| 7593045 | Ikeda et al. | Sep 2009 | B2 |
| 20010048802 | Nakajima et al. | Dec 2001 | A1 |
| 20020019833 | Hanamoto | Feb 2002 | A1 |
| 20060061598 | Mino et al. | Mar 2006 | A1 |
| 20060279555 | Ono | Dec 2006 | A1 |
| 20070168315 | Covannon et al. | Jul 2007 | A1 |
| 20080028298 | Kaneko | Jan 2008 | A1 |
| 20100223128 | Dukellis et al. | Sep 2010 | A1 |
| 20110029635 | Shkurko et al. | Feb 2011 | A1 |
| Number | Date | Country |
|---|---|---|
| 1 895 466 | Mar 2008 | EP |
| Number | Date | Country | |
|---|---|---|---|
| 20110025883 A1 | Feb 2011 | US |
