The method according to the present invention shall be better understood from the following detailed description of exemplary embodiments and read in conjunction with the appended drawing figures, of which
FIG. 1 shows a block diagram for an example search engine architecture,
FIG. 2 how document attributes can be summarized across a result set in various ways.
FIG. 3 a semantic decomposition of text with document properties embedded in the semantic structure,
FIG. 4 an example text formatted as XML,
FIG. 5 key components in an information access system for video content,
FIG. 6 the key components in a speech-to-text system,
FIG. 7 an example output from a speech-to-text analysis of the audio track in a video,
FIG. 8 how various content sources can be made available for streaming into the context of the result view in the information access,
FIG. 9 key building blocks in the construction of intelligent video summaries,
FIG. 10 an example of typical web interfaces for search engines providing access to video content,
FIG. 11 various options for visualizing a video summary,
FIG. 12 a slideshow view of frames in a video summary,
FIG. 13 the relationship between compact video thumbnails and streaming of video summaries upon user selection,
FIG. 14 another example of typical web interfaces for search engines providing access to video content,
FIG. 15 how prior art have selected a single frame 1501 or just a truncated version 1502 of the video for previews,
FIG. 16 how key frames and associated time sequences are selected and pieced together to form a video summary together with the audio track, and
FIG. 17 how the video thumbnail and the video summary can be captured in a flash encoding and offered as interactive streaming on-demand to the end user.
Now various embodiments of the method according to the present invention shall be disclosed and discussed in some details. These embodiments include client, i.e. end users' access from personal computers, mobile phones, PDAs and game devices. It shall be understood that throughout the invention the term “video”, “video data”, or “video documents” is used to denote any digital moving picture entity and thus it can be taken to mean a streamed motion picture, a video file, a webinar or live broadcast. Further the term “time sequence” in a video is intended to means consecutive images or frames between a start offset and an end offset in the video.
The method according to the present invention shall be regarded on the background of the present developing technologies for accessing, searching, retrieving and analysing information offered for instance located in document repositories available via data communication networks such as intranets and extranets. For the purpose of accessing, searching, retrieving and analysing information and content repositories such as data bases or websites, search engines are commonly employed to cater for information and search needs that are impossible to satisfy by many of for instance utility programs such as e.g. browsers as used on Internet (the World Wide Web). Hence the architecture of a search engine shall initially briefly be discussed with reference to FIG. 1 which shows the typical key components of a search engine.
In FIG. 1 the search engine is shown comprising various modules 102-108. The search engine can access content repositories 101 where content can either actively be pushed into the search engine or via a data connector be pulled into the search engine. Typical repositories include databases, sources made available via ETL (Extract-Transform-Load) tools such as Informatica, any XML-formatted repository, files from file serves, files from web servers, document management systems, content management systems, email systems, communication systems, collaboration systems, and rich media such as audio, images and video. The documents are submitted to the search engine via a content API (Application Programming Interface) 102. Subsequently, documents are analyzed in a content analysis stage 103 in order to prepare the content for improved search and discovery operations. Typically, the output of this stage is an XML representation of the input document. The output of the content analysis is used to feed the core search engine 104. The core search engine 104 can typically be deployed across a farm of servers in a distributed manner in order to allow for large sets of documents and high query loads to be processed. The core search engine 104 can accept user requests and produce lists of matching documents. The document ordering is usually determined according to a relevance model that measures the likely importance of a given document relative to the query. In addition, the core search engine 104 can produce additional metadata about the result set such as summary information for document attributes. Alternatively, the output of the content analysis stage 103 can be fed into an alert engine 105. The alert engine 105 will have stored a set of queries and can determine which queries that would have accepted the given document input. A search engine can be accessed from many different clients or applications of which are shown mobile and computer-based client applications 109; 110. Other clients include PDAs and game devices. These clients will submit requests to a search engine query API 108. The search engine will typically have means to analyze and refine the query 106 in order to construct a derived query that can extract more meaningful information. Finally, the output from the core search engine 104 is typically further analyzed in a result analysis stage in order to produce information or visualizations that are used by the clients 109 and 110.—Both stages 100 and 107 are connected between the core search engine 104 and in case also the alert engines 105, and the query API 108.
FIG. 2 illustrates how an example information access system computes document summary information. The query is first evaluated and returning a set of N matching documents. The relevance model is then used to rank the documents from most valuable to least valuable relative to the query (D1>D2> . . . DN). Typically, a given number of the highest ranking documents is returned to the client and used to construct a result page. The attributes associated with the documents can be both metadata from the original content repository or metadata that has been discovered inside the content analysis 103. The summary information is either computed based on a certain number of the highest ranking documents or on all N documents that satisfy the query. FIG. 2 illustrates some examples of summary information such as frequency, histograms, and statistical parameters. With video as the content repositories navigation can be applied to embedded structured metadata such as video size, video length, video quality, video format or associated textual metadata such as text and anchor-text on web pages associated with the video. Furthermore, text output from speech-to-text analysis can be used as primary text or input to content refinement that identifies entities used for discovery. Extracted entities from speech-to-text output can be related back to time offsets in the video. Hence, search and navigation can be related back to frames or time sequences inside the original video.
FIG. 3 illustrates contextual document analysis in an example search engine. An input document 301 is decomposed into paragraphs 306, which are further decomposed into sentences 304. Inside the sentences, ranges are marked up as document properties with specific values 302 and 303. The document properties can be recorded in both a full/original form and an associated base form. FIG. 3 illustrates two examples of this. 305 shows an example where a date is both recorded in the original form (March 2002) and a normalized base form (2002-03-XX). The normalized base form simplifies profiling, range searching and robust generation of summary information. Examples of document properties that can be recognized in a similar way include person names, company names, locations, acronyms, dates, file names, URLs, phone numbers, zip codes, email addresses, universities, newspapers, price information, product information, numerical attributes, and stock tickers. The document properties can either be associated with a given location in the document, an interval or a range in the document, a time or time interval in audio and video or a structure in an XML representation of the document. Once again, by deriving a text document from transcription of the audio in the video contextual document refinement can be used to analyze the video sequence and associate various contexts with still frames or time sequences inside the video.
FIG. 4 illustrates a typical XML that be the output of speech-to-text analysis and subsequent structuring. The document is represented in a hierarchical manner, and textual as well as any other type of attributes can be encoded at an arbitrary level. XPath and XQuery are examples of structured query languages that can be used to identify documents and sub-document structures with specific properties. In this example it is for instance possible to look for speeches where the speaker is Cleopatra and there is a line mentioning Caesar. Identified XML fragments will have associated timings and time intervals in the video sequence that were used to extract the underlying text.
FIG. 5 shows some key components of an information access system for video data. A search 501 extracts either complete videos 502 or query sequences from videos. Associated textual and structured metadata can be used to construct navigations means 503. By either selecting specific results 502 or navigating 504 the user can activate playback of the associated video 506.
FIG. 6 illustrates typical components in a speech-to-text system. Audio or the audio track inside a video is analyzed. The output data can in addition to the transcribed text contain metadata such as confidence levels, gender detection, scene changes, speaker changes etc. This output is usually represented as an XML document as described in FIG. 5 that can be further analyzed and refined with textual content refinement methods. Each word, XML tag or attribute will hence have a corresponding time offset in the associated video. Analysis of the produced XML can hence be used as a basis to identify frames or specific time sequences inside the video during both search and navigation operations in the information access system.
Phonetic transcription is an alternative to speech-to-text transcription of the audio track in the video. Phonetic transcription identifies phonemes as opposed to complete text transcriptions. The phoneme sequences are matched with a phonetic representation of the query in order to find locations inside the audio track with the best phonetic similarity. Precision is usually improved in phonetic-based search systems, but less analysis including metadata generation can be performed.
FIG. 8 illustrates how the user of the information access system can interact with streaming services for various content types.
FIG. 9 illustrates key steps in the computation of video summaries. Video summaries may be created as a selection from the original video. In order to do frame-by-frame analysis of the video, the input video must be decompressed and converted into a sequence of individual still frames. Typically each frame is represented as an image with RGB-encoded pixel values during the analytical steps in the computation of the video summary. Each frame also has an associate time delay relative to the beginning of the video and speech transcription can be related back to the same time delays from the beginning of the video. After selection of the appropriate frames for the video summary and potential resizing to client devices, the resulting frames are compressed by encoding with a video codec. Video codecs have adaptable quality/bitrate settings. Hence, by combining codec selection and extraction richness for the video summary one can adapt size and bandwidth requirements for the video summary. Multiple video summaries can either be precomputed or computed on-the-fly to match the bandwidth capabilities of the connection to the client performing the search request. The simplest selection of frames for the video summary is a section from the beginning of the video. Another way is by analyzing the video, and identifying scenes (uninterrupted camera shots), selecting an image to represent each scene. These images may then be displayed as a slideshow to quickly visualize the video content. Alternatively, a number of frames can be selected from each scene and assembled as a video. It can then be displayed as a video of short scene clips, much in the same manner as a movie trailer. These scene clips can either use every frame for a certain time interval in the original video or reduce the video rate in order to save bandwidth during the streaming process. The video codec in the video summary can also be selected independent of the video codec used in the original video. Furthermore, as some videos may have a large number of scenes, a method of selection can be used to extract the most relevant scenes.
Informative scenes can be identified and an image or sequence of images used to represent the scenes by taking recourse to several techniques. To identify scenes it is necessary to look for the interframe changes in the video, where large changes often indicate scene changes. Also, the location within the image where changes occur must be considered, as change in the background is more likely to indicate scene changes. Inter-frame differences can be calculated by looking at colour distribution, texture distribution, and object movement patterns, in addition to spatial information. The motion analysis can also selectively focus on background or foreground objects in order to separate camera-panning operations from object movements inside the scene.
After identifying the scenes in a video, the scenes and images to be used in the video summary have to be selected. By looking at image dynamics, such as the distribution of colours (i.e. how many colours) in the images, as well as texture distribution, one can avoid including images such as one-coloured black or white images, and info/title-frames (i.e. beginning and end of movies), in the video summary. Selection of which scenes to include in the video summary can be done by looking at the length and motion of a scene. To maximize the dynamics and user experience of viewing video summaries, a scene should contain a reasonable degree of motion. Also, the longest scenes are often the most important ones in a video. By combining these two criteria, the scenes can be selected that provide the user with a high quality and quick review of the complete video. In the final video summary, the selected scenes should be organized in the same temporal order as in the original video.
The selected scenes or frame sequences from the video can also be made query specific if an audio track is available. The audio track can in this case be converted to text by speech-to-text analysis and text segments can be identified by correlating the query term to the transcribed text. The identified text segments will correspond to still frames or time sequences inside the video that hence can be selected for inclusion in the video summary. The video summary could contain sufficient video frames around the occurrence of query terms to enable the user to understand the context for the occurrence inside the video.
Video thumbnails can be computed based on the same techniques as discussed above for video summaries. The thumbnails need to be loaded on the result page when the result page is loading initially. Hence, size of video thumbnails will directly impact load time and usability of the information access systems. The video thumbnail will hence typically extract much less data from the video than the video summary. Typically, just a single still frame or a small set of still frames that can be viewed in an animated way—for instance by encoding the selected frames inside an animated GIF file format.
The video thumbnails and video summaries can be constructed device independent or optimized to leverage for instance screen resolution of mobile client devices. Different devices such as e.g. personal computers, pocket PCs and mobile phones have different capabilities with regard to processing, storage, display, and network bandwidth. Different device-customized media content can be generated, as well as in real-time convert media content to meet the requirements of specific devices.
The present invention can be presented as an application such as in a PC or mobile hyperlink browser, such as for the World Wide Web. The presentation of video documents can become active when selecting the video by starting to stream i.e. audio, video, image slideshows, video summaries, or a combination of the afore-mentioned media types. This content may be represented as a sequence of different media, which are streamed upon need. This will remove unnecessary bandwidth load from inactive applications. Activation of a video item may also trigger effects to emphasize the selection. Such visual effects include, but are not limited to, changes video/image colours, video frame rate, audio volume, tint factor, and image/video display size. The user activation of a specific video result can for instance be done by hoovering, mouse over, scrolling or clicking on a video in the result set. Many technologies are available to support the actual streaming of the computed video summaries. One example is flash with embedded video in a codec such as MPEG4 that is appropriate to minimize necessary bandwidth and still maintain good viewing quality. The playback can also be done at a different speed than the original video. A slow motion display of the video thumbnails is one option. After streaming, the client can catch the streamed video summary such that repeated playbacks can be achieved without initiating additional streaming operations. The combination of minimally sized video thumbnails and high-quality video summaries streamed in context and upon request provides a very efficient environment for quickly exploring video content. Result pages will load very quickly due to the small size of the initially downloaded video thumbnail. Quick download time improves efficiency and user satisfaction of information access systems. At the same time, the user can explore the real content of the videos by streaming video summaries. Playback of video summaries are hence initiated immediately after selection by the user and the user can at any time change focus to a different video and switch the streaming process to a different source. The combination of smart bandwidth usage and live streaming also makes the present invention a very attractive approach to browsing video content from mobile devices such as cell phones, PDAs or game devices. From mobile devices a result page might be constructed as a scrollable sequence of full size video thumbnails/video summaries in order to leverage the screen size for viewing in the best possible way.
Advertisement media can be added to video summaries. This advertisement may be implemented as image frames or video sequences that cannot be skipped, before, during or after the requested media. The selected advertisements can either be pre-selected and embedded in the video summaries or dynamically integrated in the video summaries during the streaming process. By dynamically integrating advertisement media the advertisements can be selected at query time and hence follow existing models for search-based advertising such as the sponsored listings programs offered by web search providers (for instance Google and Yahoo). In addition to insertion of advertising media objects in the video summary actionable advertising elements can be inserted. Examples of actionable advertising elements include:
- HTML fragments with hyperlinks that can be accessed within the streamed view of the video summary. Hence, sponsored links as they are appearing in web search engines can be directly embedded in the video summary and enable user activation to jump to the related sponsor page.
- Actionable widgets such as buttons, links inside for instance flash that can be activated at any time by the user during viewing/streaming of the video summary.
- Display of candidate multimedia made available such that the user both gets an offer to buy and can perform the buying transaction while viewing a video summary. Multimedia examples include song or video versions of the video corresponding to the viewed video summary and related songs, videos that might be interesting based on the selection of the video summary.
In order to provide at a glance exemplification of results achievable with the method of the present invention when applied to the computation intelligent video thumbnails with rapid review and browsing capabilities within a search engine result presentation. The following figures are largely self-explanatory and hence no detailed discussion thereof will be given.
FIGS. 10 and 14 illustrate examples of typical prior art web interfaces for search engines providing access to video content.
FIG. 11 illustrates various prior art options, for visualizing a video summary.
FIG. 12 illustrates a prior art slideshow view of frames in a video summary.
FIG. 13 illustrates the relationship between compact video thumbnails and streaming of video summaries upon user selection as shall be possible with the present invention.
FIG. 15 illustrates how prior art have selected a single frame 1501 or just a truncated version 1502 of the video for previews.
FIG. 16 illustrates how key frames and associated time sequences are selected and pieced together to form a video summary together with the audio track, as shall be possible with the present invention.
FIG. 17 illustrates how the video thumbnail and the video summary can be captured in a flash encoding and offered as interactive streaming on-demand to the end user as shall be possible with the present invention.
From the foregoing it will be apparent to a person skilled in the art that the present invention greatly enhances the possibilities of displaying and visualizing video data within result presentations in information access and search systems.
It is obviously apparent that dynamic representation of video content in addition to textual descriptions can provide more informative query results to the end user. Such representation may include video summary/video thumbnail, image slideshows, audio image selection or any combinations thereof, but need by no means be limited thereto.
While the present invention has been described with reference to certain embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted without departing from the scope thereto. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the present invention without departing from its scope. Therefore, it is intended that the present invention shall not be limited to the particular embodiment disclosed, but will include all embodiments falling within the scope of the herewith appended claims.
Patent Application
20080097970
