Patent Application
20090285456
1. Field of the Invention
The present invention is a method and system to provide an automatic measurement of people's responses to visual stimulus, based on their facial expressions.
2. Background of the Invention
The human brain constantly processes sensory input to extract useful information and analyze it to understand and properly react. The expressed response is also a very important channel for revealing the internal state or for communication with other individuals, regardless of whether or not the response is voluntary. Understanding the internal psychological state of a human based on the external manifestation has been a subject of study both for scientific and practical importance.
Especially, the current consumer and market-oriented economy put a great deal of importance on people's opinions or responses to various visual stimuli—products, advertisements, or media. Most of the consumers' exposures to such visual stimuli occur in public places or retail spaces at an immeasurably high number and frequency. The ability to capture such occurrences and take measurements of the responses would provide very valuable information to retailers, marketers, or media content providers. Though it is nearly impossible to accurately determine a person's emotional response without directly asking about it, a person usually reveals some indications about it through information channels such as facial expressions or bodily gestures. It is usually the expression on the face that has high correlation with the emotional response.
Recent developments in computer vision and artificial intelligence technology make it possible to detect and track people's behavior from video sequences for further behavioral analysis. Facial image analysis has especially matured, so that faces can be detected and tracked from video images, and the motion of the facial features can also be estimated. The facial appearance changes due to facial expression can be measured to estimate the internal emotional state of a person. A host of academic studies have attempted to recognize the apparent facial expressions or emotional states of humans based on image analysis. These approaches assume more or less ideal imaging conditions and circumstances. The images are typically taken under staged circumstances so that people are instructed to show facial expressions. The facial image analysis and further recognition of expression or emotion can be carried out based on these assumptions. The proposed invention aims to solve the problem under realistic scenarios, where people show natural behavior toward visual stimuli, such as product display, advertisement media, news, movies, etc. It does not strive to measure the response—the changes in attitude or opinion in relation to the content of the visual stimulus. Furthermore, while each instance of such measurement can be erroneous, an accumulated measurement over time will provide reliable information to assess the overall attractiveness of the visual source.
While it is not entirely possible to estimate the mental state of a person based solely on the apparent changes in behavior, the changes in a person's facial expression due to the visual stimulus carry highly relevant information about a person's response through a series of both mental and physical processes. The mental state will trigger the central nervous system to transmit the information to facial muscles. The facial muscle movement then changes the shape of the facial skin so that permanent facial features change shape, or transient features such as facial wrinkles appear or change. These physical changes will render visual changes so that any capable agent (a person or a computing machine) can perceive them. People's faces can appear in any part of the scene with unknown sizes and poses, and their positions, sizes, or poses change over time. The images are also subject to varied lighting conditions. The proposed invention is designed based on this model of information channel from the internal state to the appearance changes; the steps are laid out to solve each reverse problem.
This invention adopts a series of both well-established and novel approaches for facial image processing and analysis to solve these tasks. Face detection and tracking handle the problem of locating faces and making correspondences between detected faces that belong to the same person. Face localization will normalize the facial geometry so that facial features are aligned to standard positions. Under realistic imaging conditions, the extraction of exact facial feature contours can be noisy and erroneous. This invention introduces a novel approach to extract facial appearance changes due to facial expressions; a collection of image gradient filters are designed that match the shapes of facial features or transient features. A filter that spans the whole size of the feature shape does a more robust job of extracting shapes than do local edge detectors, and will especially help to detect weak and fine contours of the wrinkles (transient features) that may otherwise be missed using traditional methods. The set of filters are applied to the aligned facial images, and the emotion-sensitive features are extracted. These features train a learning machine to find the mapping from the appearance changes to facial muscle actions. In an exemplary embodiment, the 32 Action Units from the well-known Facial Action Coding System (FACS, by Ekman & Friesen) are employed. The recognized facial actions can be translated into six emotion categories: Happiness, Sadness, Surprise, Anger, Disgust, and Fear. These categories are known to reflect more fundamental affective states of the mind: Arousal, Valence, and Stance. This invention assumes that these affective states, if estimated, provide information more directly relevant (than do the six emotion categories) to the recognition of people's attitude toward a visual stimulus. For example, the degree of valence directly reveals the positive or negative attitude toward the visual stimulus. The changes in emotional state will then render a trajectory in the three-dimensional affect space. Another novel feature of the invention is to find a mapping from the emotion trajectories to the final response. The central motivation behind this approach is that, while the emotion trajectory already contains very useful information regarding the response of the person to the visual stimulus, there can be still another level of mental process to make a final judgment, such as purchase, opinion, rating, etc. These are the kind of action that ultimately interest the marketers or content providers, and we refer to such process as ‘response’. The emotional trajectory also needs to be interpreted in the context of the dynamics of the visual stimulus. The mapping from the emotion trajectory to the response can also be estimated by training a learning machine using many samples of video sequence along with the ground-truth response data.
There have been prior attempts for detecting and localizing facial features from facial images for the purpose of further facial image analysis.
U.S. Pat. No. 5,781,650 of Lobo, et al. (hereinafter Lobo) disclosed a method for automatically finding facial images of a human face in a digital image, and classifying the age of the person into an age category. Step 1 of the process is to find facial features of the digital image encompassing the chin, sides of the face, and the virtual top of the head, eyes, mouth and nose of the image. Step 2 is to compute the facial feature ratios of the facial features found in Step 1. Step 3 is to compute a wrinkle analysis of the image. Step 4 is to combine the previous two steps to categorize the age of the facial image. The invention can locate and detect facial images for age classification from digital camera images and computerized generated images.
U.S. Pat. No. 5,852,669 of Eleftheriadis, et al. (hereinafter Eleftheriadis) disclosed a method that responds to a video signal representing a succession of frames, where at least one of the frames corresponds to an image of an object, to detect at least a region of the object. The method includes a processor for processing the video signal to detect at least the region of the object characterized by at least a portion of a closed curve and to generate a plurality of parameters associated with the closed curve for use in coding the video signal.
U.S. Pat. No. 6,219,639 of Bakis, et al. (hereinafter Bakis) disclosed a method for recognizing an individual based on attributes associated with the individual, comprising the steps of: pre-storing at least two distinctive attributes of the individual during at least one enrollment session; contemporaneously extracting the at least two distinctive attributes from the individual during a common recognition session; segmenting the pre-stored attributes and the extracted attributes according to a sequence of segmentation units; indexing the segmented pre-stored and extracted attributes so that the segmented pre-stored and extracted attributes corresponding to an identical segmentation unit in the sequence of segmentation units are associated to an identical index; and respectively comparing the segmented pre-stored and extracted attributes associated to the identical index to each other to recognize the individual.
U.S. Pat. No. 7,058,209 of Chen, et al. (hereinafter Chen) disclosed a digital image processing method that detects facial features in a digital image. This method includes the steps of detecting iris pixels in the image, clustering the iris pixels, and selecting at least one of the following schemes to identify eye positions: applying geometric reasoning to detect eye positions using the iris pixel clusters; applying a summation of squared difference method using the iris pixel clusters to detect eye positions; and applying a summation of squared difference method to detect eye positions from the pixels in the image. The method applied to identify eye positions is selected on the basis of the number of iris pixel clusters, and the facial features are located using the identified eye positions.
U.S. Pat. Appl. Pub. No. 2005/0041867 of Loy, et al. (hereinafter Loy) disclosed a method of utilizing a computer system to automatically detect the location of a face within a series of images, the method comprising the steps of: detecting eye like regions within the series of images; utilizing the eye like regions to extract potential face regions within the series of images; enhancing the facial features in the extracted potential face regions; classifying the features; and verifying the face topology within the potential face regions.
U.S. patent application Ser. No. 12/079,276 of Moon, et al. (hereinafter Moon) disclosed a method and system to provide a face-based automatic gender recognition system that utilizes localized facial features and hairstyles of humans. Given a human face detected from a face detector, it is accurately localized to facilitate the facial/hair feature detection and localization. Facial features are more finely localized using the geometrically distributed learning machines. Then the position, size, and appearance information of the facial features are extracted. The facial feature localization essentially decouples geometric and appearance information about facial features, so that a more explicit comparison can be made at the recognition stage. The hairstyle features that possess useful gender information are also extracted based on the hair region segmented, using the color discriminant analysis and the estimated geometry of the face. The gender-sensitive feature vector, made up from the extracted facial and hairstyle features, is fed to the gender recognition machines that have been trained using the same kind of gender-sensitive feature vectors of gallery images.
In Lobo, the facial feature detection is performed under close range high-resolution frontal face images to extract features for age classification. In Eleftheriadis, the facial feature detection is used for image compression, by employing edge and model-based scheme. In Bakis, the lip contour registration is performed for the purpose of multi-modal speaker recognition or verification. In Chen, eyes are detected and localized in a human face, based on the iris color signature and the cluster analysis of the iris color pixels. In Loy, eye candidates are detected first using geometric model of eye images. Based on the eye candidate locations, the facial region is detected, and other facial regions are detected and verified using geometric reasoning (facial features topology). In Moon, a combination of face localization and facial feature localization, based on training multiple learning machines on a large number of data, is used to extract features for recognizing gender.
In most of the mentioned prior inventions, either high-resolution facial images or good quality color facial images are required to reliably detect facial features. The success of these approaches also depends on successful face detection or initial (mostly eyes) features detection. In the proposed invention, an approach similar to Moon is used; the robust facial localization based on a large number of samples is performed after machine learning-based face detection. The facial features are accurately localized within already roughly localized facial feature windows, again using learning machines trained to localize only each given facial feature. The present method does not require high-resolution images or color information; it works with either gray-level or color images, and it works under various imaging conditions due to the training with a large number of images taken under various imaging conditions.
There have been prior attempts for automatically recognizing facial expression of a person using video images.
U.S. Pat. No. 5,774,591 of Black, et al. (hereinafter Black) disclosed a system that tracks human head and facial features over time by analyzing a sequence of images. The system provides descriptions of motion of both head and facial features between two image frames. These descriptions of motion are further analyzed by the system to recognize facial movement and expression. The system analyzes motion between two images using parameterized models of image motion. Initially, a first image in a sequence of images is segmented into a face region and a plurality of facial feature regions. A planar model is used to recover motion parameters that estimate motion between the segmented face region in the first image and a second image in the sequence of images. The second image is warped or shifted back towards the first image using the estimated motion parameters of the planar model, in order to model the facial features relative to the first image. An affine model and an affine model with curvature are used to recover motion parameters that estimate the image motion between the segmented facial feature regions and the warped second image. The recovered motion parameters of the facial feature regions represent the relative motions of the facial features between the first image and the warped image. The face region in the second image is tracked using the recovered motion parameters of the face region. The facial feature regions in the second image are tracked using both the recovered motion parameters for the face region and the motion parameters for the facial feature regions. The parameters describing the motion of the face and facial features are filtered to derive mid-level predicates that define facial gestures occurring between the two images. These mid-level predicates are evaluated over time to determine facial expression and gestures occurring in the image sequence.
U.S. Pat. No. 6,072,496 of Guenter, et al. (hereinafter Guenter) disclosed a method that captures a 3D model of a face, which includes a 3D mesh and a series of deformations of the mesh that define changes in position of the mesh over time (e.g., for each frame). The method also builds a texture map associated with each frame in an animation sequence. The method achieves significant advantages by using markers on an actor's face to track motion of the face over time and to establish a relationship between the 3D model and texture. Specifically, videos of an actor's face with markers are captured from multiple cameras. Stereo matching is used to derive 3D locations of the markers in each frame. A 3D scan is also performed on the actor's face with the markers to produce an initial mesh with markers. The markers from the 3D scan are matched with the 3D locations of the markers in each frame from the stereo matching process. The method determines how the position of the mesh changes from frame-to-frame by matching the 3D locations of the markers from one frame to the next. The method derives textures for each frame by removing the dots from the video data, finding a mapping between texture space and the 3D space of the mesh, and combining the camera views for each frame into a signal texture map.
U.S. Pat. No. 6,879,709 of Tian, et al. (hereinafter Tian-1) disclosed a system and method for automatically detecting neutral expressionless faces in digital images and video. First a face detector is used to detect the pose and position of a face and find the facial components. Second, the detected face is normalized to a standard size face. Then a set of geometrical facial features and three histograms in zones of mouth are extracted. Finally, by feeding these features to a classifier, the system detects if there is the neutral expressionless face or not.
U.S. patent application Ser. No. 11/685,552 of Lemos (hereinafter Lemos) disclosed a system and method for determining visual attention, and supports the eye tracking measurements with other physiological signal measurements like emotions. The system and method of the invention is capable of registering stimulus-related emotions from eye-tracking data. An eye tracking device of the system and other sensors collect eye properties and/or other physiological properties which allows a subject's emotional and visual attention to be observed and analyzed in relation to stimuli.
“Measuring facial expressions by computer image analysis,” Psychophysiology, vol. 36, issue 2, by Barlett, et al. (hereinafter Barlett) disclosed a method for facial expressions recognition that applies computer image analysis to the problem of automatically detecting facial actions in sequences of images. Three approaches were compared: holistic spatial analysis, explicit measurement of features such as wrinkles, and estimation of motion flow fields. The three methods were combined in a hybrid system that classified six upper facial actions with 91% accuracy.
“Recognizing Action Units for Facial Expression Analysis,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, no. 2, by Tian, et al. (hereinafter Tian-2) disclosed an automatic face analysis system to analyze facial expressions based on both permanent facial features (brows, eyes, mouth) and transient facial features (deepening of facial furrows) in a nearly frontal-view face image sequence. The system recognizes fine-grained changes in facial expression into Action Units (AUs) of the Facial Action Coding System (FACS), instead of a few prototypic expressions. Multistate face and facial component models are proposed for tracking and modeling the various facial features. During tracking, detailed parametric descriptions of the facial features are extracted. With these parameters as the inputs, a group of Action Units (neutral expression, six upper face AUs and 10 lower face AUs) are recognized whether they occur alone or in combinations.
“Active and dynamic information fusion for facial expression understanding from image sequences,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol 27, Issue 5, by Zhang, et al. (hereinafter Zhang) disclosed a method that uses multisensory information fusion technique with dynamic Bayesian networks (DBN) for modeling and understanding the temporal behaviors of facial expressions in image sequences. Facial feature detection and tracking based on active IR illumination provides reliable visual information under variable lighting and head motion. The approach to facial expression recognition lies in the proposed dynamic and probabilistic framework based on combining DBN with Ekman's Facial Action Coding System (FACS) for systematically modeling the dynamic and stochastic behaviors of spontaneous facial expressions. The method also actively selects the most informative visual cues from the available information sources to minimize the ambiguity in recognition. The recognition of facial expressions is accomplished by fusing not only from the current visual observations, but also from the previous visual evidences.
“Recognition of facial expressions and measurement of levels of interest from video,” IEEE Transactions on Multimedia, Volume 8, Issue 3, by Yeasin, et al. (hereinafter Yeasin) disclosed a spatio-temporal approach in recognizing six universal facial expressions from visual data and using them to compute levels of interest. The classification approach relies on a two-step strategy on the top of projected facial motion vectors obtained from video sequences of facial expressions. First a linear classification bank was applied on projected optical flow vectors and decisions made by the linear classifiers were coalesced to produce a characteristic signature for each universal facial expression. The signatures thus computed from the training data set were used to train discrete hidden Markov models (HMMs) to learn the underlying model for each facial expression. Recognized facial expressions were mapped to levels of interest using the affect space and the intensity of motion around apex frame.
In Black, the motions of the facial features due to expression are estimated by computing an explicit parametric model of optical flow. The facial feature motions are translated into mid-level predicates, which in turn are used to determine the expression categories. The proposed invention utilizes emotion-sensitive features that extract feature shape changes implicitly, just to be fed to a learning machine to estimate the facial muscle action. In Guenter, the facial actions are estimated in terms of very involved three-dimensional mesh model by tracking a set of dedicated marker points. The present invention strives to estimate the shape change of the facial features just enough to determine the facial muscle action, without using any artificial markers. Tian-1 only aims to detect emotionless faces, while the present invention tries to estimate the change of expressions in a space representing the whole range of human emotions. In Lemos, mostly eye tracking estimates are used to assess the degree of attention and the location of attention within the visual stimulus. The present invention shares a similar goal of estimating human response in relation to a given visual stimulus, but introduces a different focus on the measurement of whole facial feature shapes to determine the emotional changes to a visual stimulus, with specific technical methods to estimate the facial actions, emotional changes, and finally the response. Barlett aims to estimate upper facial Action Units utilizing the holistic, feature-based, and motion (flow)-based image representation and a neural network based learning of the representation. Tian-2 also estimates parametric models of facial feature shapes, and employs neural networks to learn the mapping to the facial Action Units. The present invention also estimates the facial Action Units in an exemplary embodiment of facial muscle actions, and utilizes learning machine to find a mapping from the image representation to the muscle actions. However, the present invention utilizes emotion-sensitive feature extraction scheme, which is different from Barlett or Tian-2. The present invention also utilizes a novel scheme to localize a face and its facial features, while in Barlett the faces are assumed to be aligned. In Zhang, the dynamic change of facial expressions is recognized by a series of methods starting from IR-based eye detection, and facial feature detection based on the eye detection. The facial Action Units recognition is based on deterministic correspondence. The present invention employs novel combination of the face detection, localization, and facial feature localization. The mapping from the facial features shapes to the facial muscle actions is learned by training on a large number of samples.
There have been prior attempts for automatically measuring the audience response to displayed objects or media.
U.S. Pat. No. 7,113,916 of Hill (hereinafter Hill) disclosed a method of assessing consumer reaction to a marketing stimulus, involving the steps of (a) exposing a sample population to a marketing stimulus for a period of time, (b) interviewing members of the sample population immediately after exposure of the members to the marketing stimulus, (c) videotaping any facial expressions and associated verbal comments of individual members of the sample population during the exposure period and interview, (d) reviewing the videotaped facial expressions and associated verbal comments of individual members of the sample population to (1) detect the occurrence of Action Units, (2) detect the occurrence of a smile, (3) categorize any detected smile as duchenne or social smile, (4) detect the occurrence of any verbal comment associated with a detected smile, and (5) categorize any associated verbal comment as positive, neutral or negative, (e) coding a single Action Unit or combination of Action Units to a coded unit, (f) associating coded units.
U.S. Pat. No. 7,120,880 of Dryer, et al. (hereinafter Dryer) disclosed a system and method for unobtrusively detecting a subject's level of interest in media content, that includes detecting to what a subject is attending, measuring a subject's relative arousal level; and combining information regarding the subject's arousal level and attention to infer a level of interest.
U.S. Pat. No. 7,233,684 of Fedorovskaya, et al. (hereinafter Fedorovskaya-1) disclosed imaging methods and systems that use affective information. In one aspect an image of a scene is captured. Affective information is collected. The affective information is associated with the image.
U.S. patent application Ser. No. 10/079,646 of Fedorovskaya, et al. (hereinafter Fedorovskaya-2) disclosed a method for collecting and associating affective information for at least one image in an imaging system that includes displaying a digital image for viewing by a particular user; automatically collecting affective information for the digital image as the particular user views the image; and associating the affective information with the particular user.
U.S. patent application Ser. No. 10/194,499 of Hazlett, et al. (hereinafter Hazlett) disclosed a method for measuring emotional and cognitive responses to advertising and other forms of communication through the use of facial electromyographic techniques.
U.S. patent application Ser. No. 11/491,411 of Thaler (hereinafter Thaler) disclosed a neural network-based rating system that includes a data set, said data set further comprising at least two records and at least one field associated with said records and a data rating application, which includes means for user input of ratings for at least a first of said records of said data set; at least one artificial neural network; means for automatically dimensioning said artificial neural network as a function of said fields within said data set; means for initiating training of said artificial neural network, said trained artificial neural network operative to generate ratings for at least a second of said records of said data set; means for initiating rating of at least said second record of said data set by said trained artificial neural network; and means for sorting said data set based on said user ratings and said artificial neural network-generated ratings.
Hill aims to measure consumer reaction to marketing stimulus, whose goal is shared by the present invention. However, Hill lists interviewing and manual video coding as tools for collecting opinions and facial expressions. Dryer proposes a system utilizing a host of measurement modalities, such as facial expression, head gesture, or speech, to assess the level of interest to media contents. Fedorovskaya-1 and Fedorovskaya-2 propose systems measuring affective information based on visual image or physiological signal of a person and associating the affective information with the image and person, respectively. Hill, Dryer, Fedorovskaya-1, and Fedorovskaya-2 all propose overall systems, without introducing a very specific novel technical means to achieve the recognition of the response or affective information. The present invention introduces novel technology to automatically extract relevant information from the raw image data and recognize the internal (mental/emotional) state of a human. Hazlett proposes an emotional response measurement based on electromyographic signal of facial muscles, while the present invention processes common visual signal to make the same kind of measurement. Thaler presents a rating system utilizing neural networks, without specific reference to how the input data to the neural network is generated. The present invention also uses learning machines such as neural networks, but the learning machines are trained to process feature vectors that are extracted from video images following novel and specific procedures.
In summary, the present invention provides fully automatic face localization and facial feature localization approaches, for accurately extracting facial and transient features to estimate facial muscle actions due to emotion changes. It is a key novel feature of the present invention to train a learning machine based on the extracted emotion-sensitive features to estimate the facial muscle action; the emotion-sensitive features are designed to extract image features that are highly correlated with the facial expressions. The present invention shares the goal of estimating human response in relation to a given visual stimulus similar to other rating approaches, but it adopts a unique method to predict the final response based on the continuous emotion trajectory, estimated over the course of the visual stimulus.
The present invention is a method and system for automatically measuring the response of a human to a visual stimulus, which utilizes emotion-sensitive features to recognize facial expressions and analyzes the emotion changes in affect space.
It is one of the objectives of the first step of the processing to train the face localization machines and the facial feature localization machines. The face localization training requires facial images having varied two-dimensional geometry—(X, Y) shifts, sizes S, and orientations O—that reflect the variations from the face detection step, along with the ground-truth values of these variations. Multiple learning machines are trained, where each machine is trained to output high response to facial images having (X, Y, S, O) close to the predetermined (X0, Y0, S0, O0) of the machine. The facial feature localization training proceeds in the same way, so that multiple learning machines are prepared and trained for each facial feature.
It is one of the objectives of the second step of the processing to detect, track, and localize faces in given images, and localize facial features. A detected and tracked face in a given image frame has a bounding box around the face that approximately delimits its position and size. The image inside the bounding box is cropped, rescaled to a standard size (for example, 30×30), and fed to the face localization machines. The face localization machines estimate the geometry in cooperation, and the facial image is localized based on these estimates. The facial feature images are cropped from the localized facial image, according to the placements of the standard feature windows. The facial feature localization machines estimate the position, size, and orientation variations of the facial features inside the standard feature windows.
It is one of the objectives of the third step of the processing to extract emotion-sensitive features. First, the step derives a set of filters that are matched to facial feature shapes or transient feature (facial wrinkles) shapes, so that the filters can extract the features relevant to facial expressions, and at the same time can ignore other image variations due to lighting and interpersonal variations, etc. A large number of emotion-sensitive feature candidate filters are generated that are designed to extract edge responses around the facial features or transient features. Then the filters are applied to many facial images showing a variety of facial expressions, and the subset of the candidate filters that gave rise to sufficient response to a large proportion of facial images are chosen as the emotion-sensitive feature filters.
Then each filter in the emotion-sensitive feature filters is applied to a correspondingly aligned feature in the image to compute the response. All the responses are collected in the emotion-sensitive feature vector.
It is one of the objectives of the fourth step of the processing to recognize the facial muscle actions. It is not straightforward to make a direct connection between the emotion-sensitive filter responses and the facial expressions due to the complex relation between the image responses and the expressions; a large number of such emotion-sensitive feature vectors along with the ground-truth expression categories are utilized to learn the relation in a machine learning framework. The trained facial muscle action recognition machine accepts the emotion-sensitive feature vector as an input and computes the likelihoods of the input face showing the corresponding muscle actions.
It is one of the objectives of the fifth step of the processing to estimate the emotion trajectory in affect space that represents the changes in emotion. The computed facial muscle actions are mapped to the six emotional categories using deterministic relations between the facial muscle actions and the six emotional categories. Then, based on the affect space coordinates of the six emotional categories, the facial muscle actions are mapped into affect space. The temporal sequence of facial muscle actions due to emotion changes constructs an emotion trajectory in affect space.
It is one of the objectives of the sixth step of the processing to interpret the emotion trajectory in affect space to derive meaningful characteristics of the response to the visual stimulus. The interpretation can be carried out in the context of the visual stimulus, so that the emotional change can be analyzed in relation to the content of the visual stimulus. The mapping from the emotion trajectory to the response can be learned by training a learning machine using many samples of video sequence along with the ground-truth response data. Exemplary response data can be ratings, purchase decisions, or expressed opinions.
In another exemplary embodiment, the visual stimulus events 990 can be recognized by the control and processing system 162, using an appropriate video segmentation algorithm, such as scene change detection.
This exemplary training scheme aims to estimate the x (horizontal) shift, y (vertical) shift, the scale, and the orientation of the right eye within the standard facial feature window 406.
The training eye images are generated by cropping the standard facial feature window 406 of the right eye from the localized face. The facial landmark points of the face are assumed to be known, and the coordinates of the landmark points 657 after going through the face localization 380 step are available.
Given an input right eye image 421, the machine having the inherent geometry of (x0, y0, s0, o0) is trained to output the likelihood of the eye image 421 having the inherent geometry. If the input training eye has the (ex, ey, es, eo), then the target output is the Gaussian likelihood: L=Exp(−(ex−x0)/kx−(ey−y0)/ky−(es−s0)/ks−(eo−o0)/ko). kx, ky, ks, and ko are constants determined empirically. (ex, ey, es, eo) can be easily determined beforehand using the coordinates of the landmark points relative to the standard facial feature positions and sizes.
Each plot in
Once each facial feature tuned machine 834 has been trained to output the likelihood of the given facial feature to have the predetermined pose vector (xi, yi, si, oi), an array of such learning machines can process any facial feature image 642 to compute the likelihoods.
A given facial feature image 642 inside the standard facial feature window 406 is fed to the trained learning machines, and then each machine outputs the responses 813 to the particular pose vector 462 (xi, yi, si, oi). The responses are then normalized 815 by dividing them by the sum of the responses to generate the weights 817. The weight is then multiplied to the corresponding pose vector (xi, yi, si, oi). The pose vectors (x1, y1, s1, o1), . . . , (xN,yN,sN,oN) are weighted and added up to compute the estimated pose vector (x*, y*, s*, o*). The pose vector represents the difference in position, scale, and orientation that the given facial feature image has against the standard feature positions and sizes. The pose vector is used to correctly extract the facial features and the transient features.
In an exemplary embodiment of the present invention, the training system 174 comprises a generic personal computer having a control and processing system 162 and an internal storage 132. A Pentium 4 2.8 GHz PC having 1 GB memory can serve as a control and processing system 162. A generic IDE hard disk drive can serve as the internal storage 132. The control and processing system applies a training algorithm to generate a trained learning machine.
In an exemplary embodiment of the present invention, the response recognition system 176 comprises the means for capturing images 100, the emotion recognition system 177, and the event detection system 178.
The means for capturing images 100 comprises a first means for capturing images 101 and a second means for capturing images. Analog cameras, USB cameras, or Firewire cameras can serve as means for capturing images.
The emotion recognition system comprises a control and processing system 162, an internal storage 132, a visual display 152, and a network connection 164. The learning machines trained from the training system 174 can be transferred to the internal storage 132 of the emotion recognition system 177 using the means for transferring data 140. The first means for capturing images 101 are connected to the control and processing system 162. The control and processing system accepts digitized video data from the first means for capturing images 101. The control and processing system 162 then processes the digitized facial images, to estimate the response. The estimated response can be stored in the internal storage 132, or can be displayed to the visual display 152, or can be transmitted remotely using the network connection 164.
In the event detection system 178, the digitized video data from the second means for capturing images 102 can be used to find meaningful events from the scene that relate to the visual stimulus. The control and processing system 162 provides the detected event information to the control and processing system 162 in the emotion recognition system 177 to help estimate the response in the emotion recognition system 178.
While the above description contains much specificity, these should not be construed as limitations on the scope of the invention, but as exemplifications of the presently preferred embodiments thereof. Many other ramifications and variations are possible within the teachings of the invention. Thus, the scope of the invention should be determined by the appended claims and their legal equivalents, and not by the examples given.
