Machine learning for recognizing and interpreting embedded information card content

Abstract

To extract metadata for highlights of a video stream from a card image embedded in the video stream.SOLUTION: Highlights may be segments of a video stream, such as a broadcast of a sporting event, that are of particular interest to one or more users. Card images embedded in video frames of the video stream are identified and processed to extract text. The text characters may be recognized by applying a machine-learned model trained with a set of characters extracted from card images embedded in sports television programming contents. A training set of character vectors may be pre-processed to maximize metric distance between the training set members. The text may be interpreted to be able to obtain the metadata. The metadata may be stored in association with a portion of the video stream. The metadata may provide information regarding the highlights, and may be presented concurrently with playback of the highlights.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] Cross - Reference to Related Applications This application claims the benefit of U.S. Provisional Patent Application No. 62 / 673,412 (Attorney Docket No. THU010 - PROV) for "Machine Learning for Recognizing and Interpreting Embedded Information Card Content", filed on May 18, 2018, the entire disclosure of which is incorporated herein by reference.

[0002] This application claims the benefit of U.S. Utility Patent Application No. 16 / 411,710 (Attorney Docket No. THU010) for "Machine Learning for Recognizing and Interpreting Embedded Information Card Content", filed on May 14, 2019, the entire disclosure of which is incorporated herein by reference.

[0003] This application claims the benefit of U.S. Provisional Patent Application No. 62 / 673,411 (Attorney Docket No. THU009 - PROV) for "Video Processing for Enabling Sports Highlights Generation", filed on May 18, 2018, the entire disclosure of which is incorporated herein by reference.

[0004] This application claims the benefit of U.S. Utility Patent Application No. 16 / 411,704 (Attorney Docket No. THU009) for "Video Processing for Enabling Sports Highlights Generation", filed on May 14, 2019, the entire disclosure of which is incorporated herein by reference.

[0005] This application claims the interests of U.S. Provisional Patent Application No. 62 / 673,413 (THU012-PROV), filed on 18 May 2018, for "Video Processing for Embedded Information Card Localization and Content Extraction," which is incorporated herein by reference in its entirety.

[0006] This application claims the interests of U.S. Utility Patent Application No. 16 / 411,713 (THU012), filed on 14 May 2019, for "Video Processing for Embedded Information Card Localization and Content Extraction," which is incorporated herein by reference in its entirety.

[0007] This application claims the interests of U.S. Provisional Patent Application No. 62 / 680,955 (THU007-PROV), filed on 5 June 2018, for "Audio Processing for Detecting Occurrences of Crowd Noise in Sporting Event Television Programming," which is incorporated herein by reference in its entirety.

[0008] This application claims the interests of U.S. Provisional Patent Application No. 62 / 712,041 (THU006-PROV), filed on 30 July 2018, for "Audio Processing for Extraction of Variable Length Disjoint Segments from Television Signal," which is incorporated herein by reference in its entirety.

[0009] This application is a "Audio Processing" application filed on October 16, 2018. for Detecting Occurrences of Loud Sound This document asserts the interest in U.S. Provisional Patent Application No. 62 / 746,454 (Agent's Application No. THU016-PROV) relating to "Characterized by Short-Time Energy Bursts," which is incorporated herein by reference in its entirety.

[0010] This application relates to U.S. Utility Patent Application No. 13 / 601,915, filed on 31 August 2012 and issued on 16 June 2015 as U.S. Patent No. 9,060,210, concerning "Generating Excitement Levels for Live Performances," which is incorporated herein by reference in its entirety.

[0011] This application relates to U.S. Utility Patent Application No. 13 / 601,927, filed on 31 August 2012 and issued on 23 September 2014 as U.S. Patent No. 8,842,007, concerning "Generating Alerts for Live Performances," which is incorporated herein by reference in its entirety.

[0012] This application relates to U.S. Utility Patent Application No. 13 / 601,933, filed on 31 August 2012 and issued on 26 November 2013 as U.S. Patent No. 8,595,763, for "Generating Teasers for Live Performances," which is incorporated herein by reference in its entirety.

[0013] This application relates to U.S. Utility Patent Application No. 14 / 510,481 (THU001), filed on 9 October 2014, for "Generating a Customized Highlight Sequence Depicting an Event," which is incorporated herein by reference in its entirety.

[0014] This application relates to U.S. Utility Patent Application No. 14 / 710,438 (THU002), filed on 12 May 2015, for "Generating a Customized Highlight Sequence Depicting Multiple Events," which is incorporated herein by reference in its entirety.

[0015] This application relates to U.S. Utility Patent Application No. 14 / 877,691 (THU004), filed on 7 October 2015, for "Customized Generation of Highlight Show with Narrative Component," which is incorporated herein by reference in its entirety.

[0016] This application relates to U.S. Utility Patent Application No. 15 / 264,928 (THU005), filed on 14 September 2016, for "User Interface for Interaction with Customized Highlight Shows," which is incorporated herein by reference in its entirety.

[0017] This paper relates to a technology that enables embedded software applications to utilize multimedia content in order to identify multimedia content and associated information on television devices or video servers that deliver multimedia content, and to provide content and services in synchronization with the delivery of multimedia content. Various embodiments relate to methods and systems for providing automated video and audio analysis used to identify and extract key event-based video segments within sports television video content, identify video highlights, and associate metadata with such highlights for pre-game, in-game, and post-game reviews. [Background technology]

[0018] Enhanced television applications such as interactive advertising and enhanced program guides, with interactive applications before, during, and after matches, have long been envisioned. Existing cable systems, originally designed for broadcast television, are required to support hosting new applications and services, including interactive television services and enhanced (interactive) program production guides.

[0019] Several frameworks have been standardized to enable augmented television applications. An example is OpenCable. (商標) Examples include the Extended TV Application Messaging Specification and the Tru2way Specification, which refer to interactive digital cable services delivered over cable video networks, including features such as interactive program guides, interactive advertising, and games. Furthermore, cable operators' "OCAP" programs offer interactive services such as e-commerce shopping, online banking, electronic program guides, and digital video recording. These efforts have enabled the first generation of video synchronization applications that are synchronized with video content delivered by program producers / broadcasters, providing additional data and interactivity to television production.

[0020] Recent advancements in video / audio content analysis technologies and compatible mobile devices have opened up a range of new possibilities in developing sophisticated applications that operate in sync with live TV program events. These new technologies, along with advancements in computer vision and video processing, and the improved computing power of modern processors, enable the real-time generation of advanced program content highlights with metadata. [Overview of the project]

[0021] A method and system for automated real-time processing of sports event television content for embedded information card localization and embedded text column recognition and interpretation is presented. In at least one embodiment, a machine learning character classification model is generated based on a training set of characters extracted from multiple information cards (card images) embedded in sports event television content. The extracted character images are processed to generate a standardized training set of multidimensional character vectors in a multidimensional vector space. Principal component analysis (PCA) is then performed on this training set to derive orthogonal basis vectors spanning the vector space of the training set.

[0022] In at least one embodiment, the dimensionality of the training set vector space is reduced by selecting a limited number of representative orthogonal vectors from an orthogonal basis. A classification model is generated for this particular set of projected alphanumeric characters appearing in the embedded information card by utilizing a machine learning algorithm structure that may be a known machine learning algorithm, such as a multiclass support vector machine (SVM) or a convolutional neural network (CNN) algorithm.

[0023] In at least one embodiment, sports event television content is processed in real time to extract queries (embedded characters from information card strings) and set up a query infrastructure using individual character images extracted from the embedded strings. In another embodiment, the individual query images are normalized to generate query vectors for each query character. These query vectors are then projected onto an orthogonal basis spanning a training vector space to generate projected query vectors. In yet another embodiment, the projected query vectors are recognized (predicted) by applying a pre-trained character classification model to each projected query vector. Finally, the predicted query characters (forming a predicted string) are interpreted by semantic extraction. In at least one embodiment, semantic extraction is performed based on known string locations in various television card image types and on knowledge of the location of individual characters within the string. In at least one embodiment, the extracted information is automatically appended to sports event metadata associated with sports event video highlights.

[0024] In at least one embodiment, a method for extracting metadata from a video stream includes storing at least one portion of the video stream, identifying one or more card images embedded in one or more video frames of the portion of the video stream, and then processing one or more information card images to extract text. In yet another embodiment, the text extracted from the information card images is interpreted to generate and store metadata in relation to the portion of the video stream.

[0025] In at least one embodiment, the video stream may be a broadcast of a sporting event. A portion of the video stream may be a highlight deemed to be of particular interest to one or more users. Metadata may describe the highlight.

[0026] In at least one embodiment, the method may further include playing a video stream for the user during at least one of identifying one or more card images, processing one or more card images, and interpreting text.

[0027] In at least one embodiment, the method may further include playing a highlight for the user and presenting metadata to the user during the playing of the highlight. The metadata may provide real-time information related to the highlight and timeline of the card image from which the metadata was obtained.

[0028] In at least one embodiment, extracting text may include identifying one or more strings within one or more card images and recording the position and / or size of the character images of the card images having one or more card images corresponding to each character of the one or more strings.

[0029] In at least one embodiment, extracting text may further include resolving ambiguity in the character boundaries of one or more strings by performing multiple comparisons of the detected character boundaries and purging character boundaries that appear too close to each other.

[0030] In at least one embodiment, extracting text may further include performing image verification on the characters of one or more strings by establishing a contrast ratio between the low-intensity pixel count and the high-intensity pixel count.

[0031] In at least one embodiment, interpreting text may include generating a query based on the text, generating an n-dimensional query feature vector, projecting the n-dimensional query feature vector onto a training set orthonormal basis, applying the projected n-dimensional query feature vector to a classification model to produce a predicted query, and extracting the meaning of the text from the predicted query.

[0032] In at least one embodiment, the method may further include generating a training set feature vector and deriving a training set orthogonal basis using the training set feature vector.

[0033] In at least one embodiment, the method may further include generating training set feature vectors and generating a classification model using the training set feature vectors and the derived training set orthogonal basis vectors.

[0034] In at least one embodiment, interpreting text may further include using at least two selections from the group consisting of the column lengths of one or more strings in the text, the character boundaries and / or character positions in the text, and the horizontal positions of character boundaries and / or characters in the text.

[0035] In at least one embodiment, storing metadata in association with a portion of a video stream may include storing the video frame numbers of one or more video frames associated with a query.

[0036] In at least one embodiment, interpreting text may include determining the field position of characters in one or more strings of text, determining the alphanumeric value of the characters, and sequentially interpreting one or more strings using the field position and the alphanumeric value.

[0037] In at least one embodiment, interpreting text may further include obtaining positional and other information relating to one or more card fields of each card image, and using the positional and other information to compensate for one or more potentially missing leading characters of one or more strings.

[0038] In at least one embodiment, a method for generating character recognition and classification models is described in relation to automatic video highlight generation. The method includes extracting and storing at least one portion of a video stream from which automatic highlight metadata is generated; identifying one or more information card images embedded in one or more video frames of the portion of the video stream; and processing one or more information card images to extract multiple character images. The method further includes generating training feature vectors associated with the multiple character images; processing the training feature vectors; training a character recognition and classification model using at least some of the training feature vectors; and then storing the processed training set and the classification model. The training feature vectors may be processed in a manner that increases the uniqueness of the training feature vectors by increasing the cross-metric distance between the training feature vectors and / or by decreasing the dimensionality of the overall vector space containing the training feature vectors.

[0039] In at least one embodiment, the method may further include normalizing the character images to a standard size and / or standard illumination before generating the training feature vectors.

[0040] In at least one embodiment, generating a training feature vector may include formatting a set of n pixels extracted from a character image into an n-dimensional vector.

[0041] In at least one embodiment, the method may further include performing principal component analysis on the training feature vectors. Training a classification model using at least some of the training feature vectors may include selecting a subset of the training feature orthogonal basis vectors and training a character recognition and classification model using the subset of the orthogonal basis vectors.

[0042] In at least one embodiment, the orthogonal basis vectors may span the entire training feature vector space. Reducing the dimensionality of the overall training feature vector space may involve selecting a limited number of orthogonal basis vectors that sufficiently accurately represent this training feature vector space. Reducing the dimensionality of the overall training vector space may involve selecting only the orthogonal basis vectors corresponding to the largest set of singular values ​​derived from the matrix of orthogonal basis vectors. Storing the classification model may involve storing a limited number of orthogonal basis vectors for subsequent use in classification model generation and / or query processing. Generating a classification model may involve using a limited number of training set orthogonal basis vectors in combination with a machine learning algorithm selected from the group consisting of SVMs and CNNs.

[0043] In at least one embodiment, the method may further include processing one or more information card images to extract text, interpreting the text to obtain metadata, and storing the metadata in association with a portion of a video stream. The method further includes playing the portion of the video stream to the user and presenting the metadata to the user while the portion of the video stream is playing. The video stream may be a broadcast of a sporting event. The portion of the video stream may include highlights that are considered to be of particular interest to one or more users. The metadata may describe the highlights.

[0044] In at least one embodiment, extracting text may include extracting a text column of text as a query.

[0045] In at least one embodiment, text extraction may include extracting at least one of the following: the current time within a sporting event, the current phase of a sporting event, a game clock related to the sporting event, and a game score related to the sporting event.

[0046] Further details and variations are described herein. [Brief explanation of the drawing]

[0047] The attached drawings, along with their descriptions, illustrate several embodiments. Those skilled in the art will recognize that the specific embodiments shown in the drawings are merely illustrative and not intended to limit the scope. [Figure 1A] This is a block diagram illustrating the hardware architecture in a client / server configuration, where event content is provided via a network-connected content provider. [Figure 1B] This is a block diagram illustrating the hardware architecture of another client / server configuration, where event content is stored in a client-based storage device. [Figure 1C] This is a block diagram illustrating the hardware architecture in a standalone embodiment. [Figure 1D] This is a block diagram illustrating the outline of a system architecture according to one embodiment. [Figure 2] This is a schematic block diagram illustrating an example of a data structure that can be incorporated into a classification model, including card images, user data, highlight data, and other elements, according to one embodiment. [Figure 3] This is a screenshot of an example video frame from a video stream, showing an information card image embedded within the frame, similar to those found in sports event television program content. [Figure 4] This flowchart illustrates an overall application process for real-time reception and processing of television program content, including the positioning of information cards within a frame, as well as content extraction and rendering, according to one embodiment. [Figure 5] This is a flowchart illustrating the internal processing of detected and extracted information card images for string bounding box extraction according to one embodiment. [Figure 6] This flowchart illustrates a method for processing text boxes for final bounded character image verification and extraction of associated position parameters, according to one embodiment. [Figure 7] This flowchart illustrates a method for generating queries from text images of embedded information cards, according to one embodiment. [Figure 8] This flowchart illustrates a method for generating predicted alphanumeric characters in a query string extracted based on a machine learning-based classification model, according to one embodiment. [Figure 9] This is a flowchart illustrating a method for interpreting a predicted query alphanumeric string according to one embodiment. [Figure 10] This flowchart illustrates the preprocessing of training set vectors and subsequent generation of a classification model based on a multi-class SVM classifier or CNN classifier, according to one embodiment. [Figure 11] This flowchart illustrates the overall process of reading and interpreting text fields within an information card and updating video highlight metadata with real-time information within the frame, according to one embodiment. [Modes for carrying out the invention]

[0048] definition The following definitions are provided for illustrative purposes only and are not intended to limit the scope. • Event: For the purposes of this description, the term “event” means a game, session, match, series, performance, program, and / or concert, or a part thereof (act, period, quarter, half, inning, scene, or chapter). An event may be a sporting event, an entertainment event, or a specific performance by a single individual or a subset of individuals within a larger group of participants in an event. Examples of non-sporting events include television shows, news flashes, socio-political events, natural disasters, films, plays, radio programs, podcasts, audiobooks, online content, and / or musical performances. An event may have any length. For illustrative purposes, this specification often describes the technology in the context of sporting events, but those skilled in the art will recognize that the technology can also be used in other contexts, including highlight shows of any audiovisual, audio, visual, graphics-based, interactive, non-interactive, or text-based content. Therefore, the use of the term “sports event” and any other sport-specific terms in this description is intended to illustrate one hypothetical embodiment, but not to limit the scope of the described technology to that one embodiment. Rather, such terms should be considered to extend to any suitable non-sports context appropriate to this technology. For ease of explanation, the term “event” is also used to refer to any report or representation of an event, such as an audiovisual recording of the event, or any other content item including a report, description, or drawing of the event. • Highlights: Excerpts or portions of an event, or content associated with an event, that are considered to be of particular interest to one or more users. Highlights can be of any length. In general, the technologies described herein provide a mechanism for identifying and presenting a customized set of highlights (which may be selected based on specific characteristics and / or user preferences) for any suitable event.The term “highlights” is also used to refer to any report or representation of highlights, such as an audiovisual recording of highlights, or any other content item that includes a report, explanation, or drawing of highlights. Highlights do not have to be limited to a drawing of the event itself, but can include other content associated with the event. For example, in the case of a sporting event, highlights could include audio / video during the match, as well as other content such as pre-match, in-match, and post-match interviews, analysis, and / or commentary. Such content can be recorded from linear television (for example, as part of a video stream depicting the event itself) or taken from any number of other sources. Various types of highlights can be provided, for example, including occurrences (plays), strings, possessions, and sequences, all of which are defined below. Highlights do not have to be of a fixed duration, but can incorporate start and / or end offsets, as described below. • Content delineator: One or more video frames that indicate the start or end of a highlight. • Occurrence: Something that happens during the event. Examples include goals, plays, downs, hits, saves, shots over the basket, baskets, steals, snaps or attempted snaps, near misses, fights, the start or end of a game, quarters, halves, periods, or innings, pitches, penalties, injuries, dramatic events in entertainment events, songs, and / or solos. Occurrences can also be unusual events, such as power outages and / or incidents with unruly fans. The detection of such occurrences can be used as a basis for determining whether to designate a particular portion of a video stream as a highlight. Occurrences are also referred to as “plays” herein for ease of naming, but such usage should not be interpreted as limiting the scope. Occurrences may have any length, and expressions of occurrences may have varying lengths.For example, as described above, an extended representation of an occurrence may include footage depicting the time period immediately before and after the occurrence, while a simple representation may include only the occurrence itself. Any intermediate representations may also be provided. In at least one embodiment, the choice of duration for representing an occurrence may vary depending on user preference, available time, determined level of excitement for the occurrence, importance of the occurrence, and / or any other factors. Offset: An amount that adjusts the length of a highlight. In at least one embodiment, a start offset and / or end offset can be provided to adjust the start time and / or end time of a highlight, respectively. For example, if a highlight depicts a goal, the highlight may be extended by a few seconds (via an end offset) to include celebration and / or fan reaction following the goal. The offset can be configured to change automatically or manually based on, for example, the available time for the highlight, the importance and / or level of excitement for the highlight, and / or any other preferred factors. String: A series of occurrences that are linked or related to each other in some way. An occurrence may occur within a possession (as defined below) or span multiple possessions. An occurrence may occur within a sequence (as defined below) or span multiple sequences. Occurrences may be linked or related because they have some thematic or narrative connection to one another, or because one leads to another, or for any other reason. An example of a string is a set of passes that lead to a goal or basket. This should not be confused with a “text string” which has the meaning usually assigned in the field of computer programming. Possession: A portion of an event that is separated by any given time. The distinction between the start and end times of a possession may vary depending on the type of event. In certain sporting events where one team may be offensive and the other is defensive (e.g., basketball or soccer), possession can be defined as the period of time when one of the teams has the ball.In sports where possession of the puck or ball is more fluid, such as hockey or soccer, possession is considered to extend to a period of time during which one team has substantial control of the puck or ball, disregarding momentary contact by the other team (such as a blocked shot or save). In baseball, possession is defined as a half-inning. In soccer, possession can include several sequences in which the same team has the ball. In other types of sporting and non-sporting events, the term "possession" may be somewhat misleading, but it is still used here for illustrative purposes. Examples in non-sporting contexts include chapters, scenes, acts, or television segments. For example, in the context of a music concert, possession may correspond to the performance of a single song. Possession can include any number of occurrences. Sequence: A time-divided portion of an event that contains a period of time of one consecutive action. For example, in a sporting event, a sequence may begin at the start of an action (such as a face-off or tip-off) and end when a whistle is blown to indicate an interruption of the action. In sports such as baseball or soccer, a sequence may be equivalent to a play, and this is a form of occurrence. A sequence may contain any number of possessions, or may be a part of a possession. • Highlight show: A set of highlights arranged for presentation to the user. A highlight show may be presented linearly (e.g., as a video stream) or in a way that allows the user to choose which highlights to watch and in what order (e.g., by clicking links or thumbnails). The presentation of a highlight show may be non-interactive or interactive, allowing the user to pause, rewind, skip, fast forward, and / or communicate preferences, etc. A highlight show may be, for example, a condensed match.A highlight show can contain any number of consecutive or discontinuous highlights from a single event or from multiple events, and may even include highlights from different types of events (e.g., different sports, and / or combinations of sports and non-sports events). • User / Audience: The terms “user” or “audience” are synonymous with any individual, group, or other entity that will see, hear, or otherwise experience an event, one or more highlights of an event, or a highlight show. The terms “user” or “audience” may also refer to any individual, group, or other entity that will see, hear, or otherwise experience any of the event, one or more highlights of an event, or a highlight show at some future point in time. While the term “audience” may be used for descriptive purposes, since an event does not need to contain a visual component, “audience” could instead be a listener or any other consumer of the content. • Narrative: A coherent story linking a set of highlight segments in a specific order. • Excitement Level: A measure of how exciting or interesting an event or highlights will be for a particular user or general audience. The excitement level may also be determined in relation to a particular occurrence or player. Various techniques for measuring or evaluating excitement levels are described in the relevant applications referenced above. As described, excitement levels may vary depending on the occurrences within an event and other factors such as the overall context or importance of the event (e.g., playoff games, pennant implications, and / or rivalries). In at least one embodiment, excitement levels can be associated with each occurrence, string, possession, or sequence within an event. For example, the excitement level of a possession can be determined based on the occurrences that occur within that possession. Excitement levels may be measured differently by different users (e.g., fans of a particular team and neutral fans) and may vary depending on each user's personal characteristics.• Metadata: Data that relates to and is stored in association with other data. Primary data may be media such as a sports program or highlights. • Card image: An image within a video frame that provides data about any of the things depicted in the video, such as an event, a depiction of an event, or a part thereof. An exemplary card image may include match scores, a match clock, and / or other statistics from a sporting event. Card images may appear temporarily or over the entire duration of the video stream, and those appearing temporarily may be particularly related to the portion of the video stream in which they appear. • Character image: A portion of an image that appears to be related to a single character. A character image may also include the area surrounding the character. Good. For example, a character image may include a roughly rectangular bounding box surrounding the character. • Character: A symbol that can be part of a word, number, or representation of a word or number. Characters can include letters, numbers, and special characters, and may exist in any language. • String: A set of characters grouped in a way that indicates they relate to a single piece of information, such as the names of teams playing in a sporting event. Often, English strings are arranged horizontally and read from left to right. However, strings may be arranged differently in English than in other languages.

[0049] overview According to various embodiments, methods and systems are provided for automatically creating time-based metadata associated with television highlights of sporting events. The highlights and associated in-frame time-based information may be extracted synchronously with respect to the television broadcast of the sporting event, or while the video content of the sporting event is being streamed from a backup device through a video server after the television broadcast of the sporting event.

[0050] In at least one embodiment, a software application operates in sync with the playback and / or reception of television program content to provide informational metadata associated with highlights of the content. Such software may run, for example, on the television device itself, or on an associated set-top box (STB), or on a video server capable of receiving and subsequently streaming program content, or on a mobile device capable of receiving video feeds, including live programs. In at least one embodiment, the highlights and associated metadata application operates in sync with the presentation of the television program content.

[0051] Interactive television applications can enable timely and relevant presentations of highlighted television program content to users watching television programs on either a primary television display or a secondary display such as a tablet, laptop, or smartphone. Sets of video clips representing highlights of television broadcast content can be generated and / or stored in real time, along with a database containing time-based metadata that provides more detailed descriptions of the events presented by the highlighted video clips.

[0052] Metadata associated with a video clip can be any information, such as text information, a set of images, and / or any type of audiovisual data. One type of metadata associated with highlights of video content during and after a match conveys real-time information about sports match parameters extracted directly from live broadcast content by reading information cards ("card images") embedded in one or more video frames of the broadcast content. In at least one embodiment, the described system and method enable the automatic generation of this type of metadata and thus associate card image content with video highlights of an analyzed digital video stream.

[0053] Various embodiments describe an automated process that includes receiving a digital video stream, analyzing one or more video frames of the digital video stream for presenting and extracting card images, locating text boxes within the card images, and recognizing and interpreting sequences of characters present within the text boxes.

[0054] The automated metadata generation video system presented herein can receive live broadcast video streams or digital video streamed via a computer server, process the video stream in real time using computer vision and machine learning techniques, and extract metadata from embedded information cards.

[0055] In at least one embodiment, a string associated with an extracted information card text field is identified, and the position and size of the image of each character within the string of characters are recorded. Subsequently, any number of characters within the text string from various fields of the information card are recognized, and the text string containing the recognized characters is interpreted to provide real-time information related to a television broadcast of a sporting event, such as the current time and phase of the match, the match score, and / or play information.

[0056] In another embodiment, individual character images are extracted from an embedded string and then used to generate normalized query vectors. These normalized query vectors are then projected onto an orthogonal basis spanning the training vector space, and these training vectors are pre-assembled and used to train a machine learning classifier, such as a multi-class support vector machine (SVM) classifier (e.g., C. BURGES, "A Tutorial on Support Vector Machines for Pattern Recognition", Kluwer Academic Publishers, 1998). The projected queries are then used to generate query predictions as the output of a pre-trained classification model produced by an exemplary SVM training mechanism. It should be noted that the classification model is not limited to an SVM-based model. The classification model may also use other techniques such as convolutional neural networks (CNNs) and numerous variations of the CNN algorithm mechanism (e.g., Y. LeCun at It can also be created using al., “Efficient NN Back Propagation”, Springer 1998), and this modification is suitable for the training dataset presented herein.

[0057] In yet another embodiment, query character prediction is generated by applying a projected query character vector to a pre-developed, machine learning-based classification model. This step generates a sequence of predicted characters according to pre-established classification labels, and the predicted sequence of alphanumeric characters is passed to a recognition and interpretation process. The query recognition and interpretation process applies prior knowledge and positional understanding of characters present in numerous information card fields. The meaning of each predicted alphanumeric character positioned within a specific character group is further interpreted, and the derived information is added to the video highlight metadata handled by the video highlight generation application.

[0058] In yet another embodiment, character classification model generation is considered, and the model is based on a training set of characters extracted from any number of information cards embedded in sports event television content. Character bounding boxes are detected, and characters are extracted from a large number of information cards. These character images are then normalized to standardized size and illumination to form descriptors associated with each particular character from the set of alphanumeric characters appearing in the embedded information cards. In this way, each extracted character image represents an n-dimensional vector in a multidimensional vector space containing a training set of vectors. The n-dimensional training vectors representing the set of character images are further processed to increase uniqueness and cross-metric distance, and to reduce the overall dimensionality of the vector space of the training vectors.

[0059] In at least one embodiment, principal component analysis (e.g., G. Golub and F. Loan, “Matrix Computations”, Johns Hopkins) is used. The method (Univ. Press, Baltimore, 1989) is performed on the training vector set. Therefore, the orthogonal basis of vectors is devised from the training set such that the orthogonal basis vectors span the training set's vector space. Furthermore, the dimensionality of the training set vector space is reduced by selecting a limited number of orthogonal basis vectors so that only the most important orthogonal vectors associated with the largest set of singular values, generated by the singular value decomposition of the training set matrix of basis vectors, are retained. The selected training set basis vectors are then saved for later use in generating a classification model using one or more of the algorithmic structures available for dataset classification, such as a multi-class SVM-based classifier or a CNN-based classifier.

[0060] System Architecture According to various embodiments, the system can be implemented in any electronic device or set of electronic devices equipped to receive, store, and present information. Such electronic devices may be, for example, desktop computers, laptop computers, televisions, smartphones, tablets, music players, audio devices, kiosks, set-top boxes (STBs), game systems, wearable devices, and / or home electronic devices.

[0061] While the system is described herein in relation to its implementation on a particular type of computing device, those skilled in the art will recognize that the techniques described herein can be implemented in other contexts and, in fact, can be implemented on any suitable device capable of receiving and / or processing user input and presenting output to the user. Therefore, the following description is intended to illustrate various embodiments as examples, rather than to limit the scope.

[0062] Referring here to Figure 1A, a block diagram is shown illustrating the hardware architecture of system 100 for automatically extracting metadata from card images embedded in a video stream of an event, in a client / server embodiment. Event content, such as a video stream, may be provided via a network-connected content provider 124. An example of such a client / server embodiment is a web-based implementation, where each of one or more client devices 106 runs a browser or application that provides a user interface for interacting with content from various servers 102, 114, 116, including data provider server 122 and / or content provider server 124, via a communication network 104. The transmission of content and / or data in response to requests from client devices 106 can be done using any known protocol and language, such as hypertext markup language (HTML), Java, Objective-C, Python, and / or JavaScript.

[0063] The client device 106 may be a desktop computer, laptop computer, television, smartphone, tablet, music player, audio device, kiosk, set-top box, game system, wearable device, home electronic device, and / or any electronic device. In at least one embodiment, the client device 106 has several hardware components known to those skilled in the art. The input device(s) 151 may be any component(s) that receives input from the user 150, and may include, for example, a handheld remote control, keyboard, mouse, stylus, touch-sensitive screen (touchscreen), touchpad, gesture receptor, trackball, accelerometer, five-way switch, or microphone. The input may be provided via any preferred mode, including, for example, one or more of pointing, tapping, typing, dragging, gestures, tilting, shaking, and / or speech. The display screen 152 may be any component that graphically displays information, video, and / or content, including drawing events and / or highlights. Such outputs may also include, for example, audiovisual content, data visualizations, navigation elements, graphic elements, or queries requesting information and / or parameters for content selection. In at least one embodiment where only some of the desired outputs are presented at a time, dynamic controls such as a scrolling mechanism may be available via the input device 151 to select which information is currently displayed and / or to change how the information is displayed.

[0064] The processor 157 may be a conventional microprocessor for performing operations on data under the direction of software, in accordance with well-known art. The memory 156 may be a random-access memory having a structure and architecture known in the art for use by the processor 157 in the process of running software for performing the operations described herein. The client device 106 may also include local storage (not shown), which may be a hard drive, flash drive, optical or magnetic storage device, and / or web-based (cloud-based) storage.

[0065] Any preferred type of communication network 104, such as the Internet, television networks, cable networks, and / or cellular networks, can be used as a mechanism for transmitting data between a client device 106 and various servers 102, 114, 116 and / or content providers 124 and / or data providers 122, in accordance with any preferred protocol and technology. In addition to the Internet, other examples include mobile phone networks, EDGE, 3G, 4G, Long Term Evolution (LTE), Session Initiation Protocol (SIP), Short Message Peer-to-Peer Protocol (SMPP), SS7, Wi-Fi, Bluetooth®, ZigBee, Hypertext Transfer Protocol (HTTP), Secure Hypertext Transfer Protocol (SHTTP), and / or Transmission Control Protocol / Internet Protocol (TCP / IP), as well as any combination thereof. In at least one embodiment, a client device 106 sends a request for data and / or content via a communication network 104 and receives a response from servers 102, 114, 116 containing the requested data and / or content.

[0066] In at least one embodiment, the system of Figure 1A operates in connection with a sporting event. However, it should be understood that the teachings herein also apply to non-sporting events, and the techniques described herein are not limited to applications to sporting events. For example, the techniques described herein can be used to operate in connection with, or for multiple such events, television shows, movies, news events, sports shows, political activities, business shows, dramas, and / or other episodic content.

[0067] In at least one embodiment, system 100 identifies highlights of a broadcast event by analyzing a video stream of the event. This analysis can be performed in real time. In at least one embodiment, system 100 includes one or more web servers 102 coupled to one or more client devices 106 via a communication network 104. The communication network 104 may be a public network, a private network, or a combination of public and private networks such as the Internet. The communication network 104 may be a LAN, WAN, wired, wireless, and / or a combination of the above. In at least one embodiment, the client devices 106 may connect to the communication network 104 via either a wired or wireless connection. In at least one embodiment, the client devices may also include a recording device capable of receiving and recording events, such as a DVR, PVR, or other media recording device. Such a recording device may be part of or external to the client device 106. In other embodiments, such a recording device may be omitted. Figure 1A shows one client device 106, but system 100 can implement any number of client devices 106 of a single type or multiple types.

[0068] The web server(s) 102 may include one or more physical computing devices and / or software capable of receiving requests from client devices(s) 106, responding to those requests with data, and sending unsolicited alerts and other messages. The web server(s) 102 may employ various strategies for fault tolerance and scalability, such as load balancing, caching, and clustering. In at least one embodiment, the web server(s) 102 may include caching techniques, as known in the art, for storing information related to client requests and events.

[0069] The web server(s) 102 may maintain or otherwise designate one or more application servers(s) 114 to respond to requests received from client devices(s) 106. In at least one embodiment, the application servers(s) 114 provide access to business logic for use by client application programs within the client devices(s) 106. The application servers(s) 114 may be located in the same location as the web server(s) 102, shared, or co-managed. The application servers(s) 114 may also be located separately from the web server(s) 102. In at least one embodiment, the application servers(s) 114 interact with one or more analysis servers(s) 116 and one or more data servers(s) 118 to perform one or more operations of the disclosed technology.

[0070] One or more storage devices 153 can function as a “data store” by storing data related to the operation of the system 100. This data may include, but is not limited to, card data 154 related to card images embedded in a video stream presenting an event such as a sporting event, user data 155 related to one or more users 150, highlight data 164 related to one or more highlights of an event, and / or a classification model 165 that can be used to predict and / or extract text from the card data 154.

[0071] The card data 154 may include any information related to the card images embedded in the video stream, such as the card images themselves, subsets thereof such as character images, text extracted from the card images such as characters and strings, and any of the aforementioned attributes that facilitate the extraction of text and / or semantics. The user data 155 may include any information describing one or more users 150, such as demographics, purchasing behavior, video stream viewing behavior, interests, and / or preferences. The highlight data 164 may include highlights, highlight identifiers, time metrics, categories, excitement levels, and other data related to the highlights. The classification model 165 may include a machine-trained classification model, queries, query feature vectors, an orthogonal base of the training set, predicted queries, the meaning of extracted text, and / or other information that facilitates the extraction of text and / or semantics from the card data 154. The card data 154, user data 155, highlight data 164, and classification model 165 will be described in detail below.

[0072] In particular, many components of system 100 may be or may include computing devices. Each such computing device may have an architecture similar to that of client device 106, as shown and described above. Thus, any one of the communication network 104, web server 102, application server 114, analysis server 116, data provider 122, content provider 124, data server 118, and storage device 153 may include one or more computing devices, which may optionally have an input device 151, a display screen 152, memory 156, and / or a processor 157, as described above in relation to client device 106.

[0073] In an exemplary operation of system 100, one or more users 150 of client device 106 view content from content provider 124 in the form of a video stream. The video stream may represent an event, such as a sporting event. The video stream may be a digital video stream that can be easily processed with known computer vision techniques.

[0074] When a video stream is displayed, one or more components of System 100, such as a client device 106, a web server 102, an application server 114, and / or an analysis server 116, may analyze the video stream, identify highlights within the video stream, and / or extract metadata from the video stream, for example, from embedded card images and / or other aspects of the video stream. This analysis can be performed in response to the receipt of a request to identify highlights and / or metadata of the video stream. Alternatively, in another embodiment, highlights can be identified without any specific request from the user 150. In yet another embodiment, the analysis of the video stream can be performed without the video stream being displayed.

[0075] In at least one embodiment, user 150 can specify specific parameters for video stream analysis via input device 151(or more) of client device 106 (e.g., which events / matches / teams to include, how much time user 150 has available to watch the highlights, what metadata is desired, and / or any other parameters). User preferences can also be extracted from storage, such as from user data 155 stored in one or more storage devices 153, to customize the video stream analysis without necessarily requiring user 150 to specify preferences. In at least one embodiment, user preferences can be determined based on observed behavior and actions of user 150, such as observing website visit patterns, television viewing patterns, music listening patterns, online purchases, prior highlight identification parameters, and / or highlights and / or metadata actually viewed by user 150.

[0076] Additionally or alternatively, user preferences can be retrieved from pre-stored preferences explicitly provided by user 150. Such user preferences can indicate which teams, sports, players, and / or event types are of interest to user 150, and / or they can indicate what types of metadata or other information related to highlights would be of interest to user 150. Thus, such preferences can be used to guide the analysis of the video stream to identify highlights and / or extract metadata for highlights.

[0077] An analysis server (or more) 116, which may include one or more computing devices as described above, can analyze live and / or recorded feeds of commentary statistics related to one or more events from a data provider (or more) 122. Examples of data providers (or more) 122 include, but are not limited to, STATSTM, Perform (available from Opta Sports, London, UK), and real-time sports information providers such as SportRadar, St. Gallen, Switzerland. In at least one embodiment, the analysis server (or more) 116 generates a set of different excitement levels for an event. Such excitement levels can then be stored in association with highlights identified by the system 100, according to the techniques described herein.

[0078] Application servers 114 may analyze the video stream to identify highlights and / or extract metadata. Additionally or alternatively, such analysis may be performed by client devices 106. Identified highlights and / or extracted metadata may be specific to user 150, in which case it may be advantageous to identify highlights within client devices 106 associated with a particular user 150. As described above, client devices 106 may receive, retain, and / or retrieve applicable user preferences for highlight identification and / or metadata extraction. Additionally or alternatively, highlight generation and / or metadata extraction may be performed globally (i.e., using objective criteria generally applicable to a user population, regardless of the preferences of a particular user 150). In such cases, it may be advantageous to identify highlights and / or extract metadata within application servers 114.

[0079] Content that facilitates highlight identification and / or metadata extraction may come from any preferred source, including content provider(s) 124, including YouTube® and websites such as MLB.com, sports data providers, television stations, and / or client or server-based DVRs. Alternatively, the content may come from a local source, such as a DVR or other recording device associated with (or embedded in) client device 106. In at least one embodiment, application server(s) 114 generates a customized highlight show with highlights and metadata available to user(s) 150, either as downloadable or streaming content, or on-demand content, or in any other way.

[0080] As described above, it may be advantageous to perform user-specific highlight identification and / or metadata extraction on a specific client device 106 associated with a specific user 150. Such an embodiment can avoid the need to unnecessarily transmit video content or other high-bandwidth content over the communication network 104, especially if such content is already available on the client device 106.

[0081] For example, referring to Figure 1B, an example of a system 160 in one embodiment is shown, in which card data 154, highlight data 164, and at least some of the classification models 165 are stored in a client-based storage device 158, which may be any form of local storage device available to the client device 106. An example would be a DVR capable of recording events, such as video content of a complete sporting event. Alternatively, the client-based storage device 158 may be any magnetic, optical, or electronic storage device for data in digital format. Examples include flash memory, magnetic hard drives, CD-ROMs, DVD-ROMs, or other devices integrated with or communicatively coupled to the client device 106. Based on information provided by the application server 114, the client device 106 may extract metadata from the card data 154 stored in the client-based storage device 158 and store that metadata as highlight data 164, without requiring the retrieval of other content from the content provider 124 or other remote source. Such a configuration can save bandwidth and make effective use of existing hardware that may already be available to the client device 106.

[0082] Returning to Figure 1A, in at least one embodiment, the application server(s) 114 can identify different highlights and / or extract different metadata for different users 150, depending on individual user preferences and / or other parameters. The identified highlights and / or extracted metadata may be presented to the user 150 via any suitable output device, such as the display screen 152 of the client device 106. If necessary, multiple highlights can be identified and assembled into a highlight show along with their associated metadata. Such a highlight show may be assembled into a “highlight reel” or set of highlights that is accessed via a menu and / or played for the user 150 according to a predetermined sequence. In at least one embodiment, the user 150 can control the highlight playback and / or delivery of associated metadata via an input device(s) 151 for purposes such as: · Selecting a specific highlight and / or metadata for display. · Pausing, rewinding, and fast-forwarding. · Skipping to the next highlight. · Returning to the beginning of the previous highlight in the highlight show. and / or · Performing other actions.

[0083] Further details regarding these features are provided in the relevant U.S. patent application cited above.

[0084] In at least one embodiment, another data server(s) 118 is provided. The data server(s) 118 may respond to requests for data from any of the servers(s) 102, 114, 116 to retrieve or provide, for example, card data 154, user data 155, highlight data 164, and / or classification models 165. In at least one embodiment, such information may be stored in any preferred storage device 153 accessible by the data server(s) 118 and may come from any preferred source, such as the client device(s) 106 itself, the content provider(s) 124, and / or the data provider(s) 122.

[0085] Referring here to Figure 1C, System 180 is shown in an alternative embodiment in which System 180 is implemented in a standalone environment. Similar to the embodiment shown in Figure 1B, at least some of the card data 154, user data 155, highlight data 164, and classification model 165 may be stored in a client-based storage device 158 such as a DVR. Alternatively, the client-based storage device 158 may be flash memory or a hard drive, or another device integrated with or communicatively coupled to the client device 106.

[0086] User data 155 may include the preferences and interests of user 150. Based on such user data 155, system 180 may extract metadata from card data 154 and present it to user 150 in the manner described herein. Additionally or alternatively, metadata may be extracted based on objective criteria that are not based on information specific to user 150.

[0087] Referring here to Figure 1D, an overview of a system 190 having an architecture according to an alternative embodiment is shown. In Figure 1D, the system 190 includes broadcast services such as content provider(s) 124, content receivers in the form of client devices 106 such as television sets with STBs, video servers such as analysis servers 116 capable of capturing and streaming television program content, and / or other client devices 106 such as mobile devices and laptops capable of receiving and processing television program content, all connected via a network such as a communication network 104. A client-based storage device 158 such as a DVR can be connected to any of the client devices 106 and / or other components and can store video streams, highlights, highlight identifiers, and / or metadata to facilitate the identification and presentation of highlights and / or extracted metadata via any of the client devices 106.

[0088] The specific hardware architectures depicted in Figures 1A, 1B, 1C, and 1D are for illustrative purposes only. Those skilled in the art will recognize that the techniques described herein can be implemented using other architectures. Many of the components depicted herein are optional and may be omitted, integrated with other components, and / or replaced by other components.

[0089] In at least one embodiment, the system may be implemented as software written in any suitable computer programming language, whether standalone or in a client / server architecture. Alternatively, it may be implemented and / or embedded in hardware.

[0090] data structure Figure 2 is a schematic block diagram illustrating an example of a data structure that can be incorporated into card data 154, user data 155, highlight data 164, and classification model 165 according to one embodiment.

[0091] As shown, the card data 154 may include a record of each of several card images embedded in one or more video streams. Each card image may include one or more strings 200. Each string 200 may have a record of n characters. Each such record may have a character image 202, a processed character image 203, a character boundary 204, a size 205, a position 206, a contrast ratio 207, and / or an interpretation 208. Each string 200 may further have a column length 209 indicating the length of the string 200 (e.g., the length of characters or pixels).

[0092] The character image 202 may be a specific portion of a card image containing a single character. The processed character image 203 may be the character image 202 after applying one or more processing steps, such as normalization of size and / or brightness.

[0093] The character boundary 204 may represent the character image 202, the processed character image 203, and / or the boundary of the character represented by the character image 202 and the processed character image 203.

[0094] Size 205 may be the size of the character image 202, the processed character image 203, and / or the character represented by the character image 202 and the processed character image 203, for example, the size of a pixel.

[0095] Position 206 may be the position of the character image 202, the processed character image 203, and / or the character represented by character image 202 and processed character image 203 within the card image. In some examples, position 206 may represent a two-dimensional position (e.g., the x and y coordinates of the corner or center of the character represented by character image 202, the processed character image 203, and / or character image 202 and processed character image 203).

[0096] The contrast ratio 207 may also be an indicator of the contrast of the character image 202, the processed character image 203, and / or the characters represented by the character image 202 and the processed character image 203. In some examples, the contrast ratio 207 may also be the ratio of the luminance values ​​of one or more brightest pixels to the luminance values ​​of one or more darkest pixels in the characters represented by the character image 202, the processed character image 203, and / or the characters represented by the character image 202 and the processed character image 203.

[0097] Interpretation 208 may be a specific character that is thought to be represented in character image 202 after some analysis has been performed to interpret string 200, such as a, b, c, 1, 2, 3, #, &, etc.

[0098] The structure of the card data 154 shown in Figure 2 is merely illustrative, and in some embodiments, the data associated with the card images embedded in the video stream may be organized differently. For example, in other embodiments, each string may not necessarily be broken down into individual character images. Rather, the string may be interpreted as a whole, and data useful for interpreting the string may be stored for the entire string. Furthermore, in alternative embodiments, data not specifically described above may be incorporated into the card data 154. The structures of the user data 155, highlight data 164, and classification model 165 in Figure 2 are also merely illustrative, and many alternatives can be conceived by those skilled in the art.

[0099] As further shown, user data 155 may include records related to user 150, each of which may include demographic data 212, preferences 214, viewing history 216, and purchase history 218 of a particular user 150.

[0100] Demographic data 212 may include, but is not limited to, any type of demographic data, including, age, sex, location, nationality, religious affiliation, and / or education level.

[0101] The preferences 214 may include choices made by the user 150 regarding their own preferences. The preferences 214 may be directly related to the collection and / or display of highlights and metadata, or they may be of a more general nature. In either case, the preferences 214 can be used to facilitate the identification and / or presentation of highlights and metadata to the user 150.

[0102] Viewing history 216 may list TV shows, video streams, highlights, web pages, search queries, sporting events, and / or other content retrieved and / or viewed by user 150.

[0103] Purchase history 218 can list products or services purchased or requested by user 150.

[0104] As further shown, the highlight data 164 may include a record of the j-highlights 220, each of which may include a video stream 222, identifier, and / or metadata 224 for a particular highlight 220.

[0105] Video stream 222 may contain a video depicting highlight 220, which may be obtained from one or more video streams of one or more events (for example, by trimming the video streams to include only video stream 222 related to highlight 220). Identifier 223 may contain a time code and / or other indicators indicating where highlight 220 is located within the video stream of the event from which it was obtained.

[0106] In some embodiments, each recording of highlight 220 may include only one of the video stream 222 and identifier 223. Highlight playback may be performed by playing the user 150's video stream 222, or by playing only the highlighted portion of the video stream of the event in which the highlight 220 is obtained using identifier 223.

[0107] Metadata 224 may include information about the highlight 220, such as the date of the event, the season, and information about the groups or individuals involved in the event or video stream from which the highlight 220 was taken, such as teams, players, coaches, anchors, broadcasters, and / or fans. Among other information, the metadata 224 for each highlight 220 may include time 225, phase 226, clock 227, score 228, and / or frame number 229.

[0108] Time 225 may be the time in video stream 222 when the highlight 220 is acquired, or the time in video stream 222 related to the highlight 220 when metadata is available. In some examples, time 225 may be the playback time in video stream 222 related to the highlight 220 when the card image containing metadata 224 is displayed.

[0109] Phase 226 may be a phase of an event related to Highlight 220. More specifically, Phase 226 may be a stage of a sports event in which a card image containing metadata 224 is displayed. For example, Phase 226 may be "Third Quarter," "Second Inning," or "Bottom Half," etc.

[0110] Clock 227 may be the match clock associated with the highlight 220. More specifically, clock 227 may be the state of the match clock when the card image containing metadata 224 is displayed. For example, clock 227 may be "15:47" in the case of a card image that displays 15 minutes and 47 seconds on the match clock.

[0111] Score 228 may be the match score associated with Highlight 220. More specifically, score 228 may be the score displayed when the card image containing metadata 224 is shown. For example, score 228 may be "45-38", "7-0", or "30-love", etc.

[0112] Frame number 229 may be the number of a video frame in the video stream from which the highlight 220 is obtained, or the number of the video frame in the video stream 222 associated with the highlight 220 that is most directly related to the highlight 220. More specifically, frame number 229 may be the number of such a video frame in which the card image containing metadata 224 is displayed.

[0113] As further shown, the classification model 165 may include various pieces of information to facilitate the extraction and interpretation of the string 200. The classification model 165 may then enable the automatic generation of metadata 224 for the highlights 220. Specifically, the classification model 165 may include a query 230, a query feature vector 232, an orthogonal basis 234, a predicted query 236, and / or the meaning of the text 238.

[0114] The behavior of query 230, query feature vector 232, orthogonal basis 234, and predicted query 236 is described in more detail herein. The text meaning 238 may be an interpretation of string 200 rendered in a manner that can be easily copied into metadata 224.

[0115] The data structure shown in Figure 2 is for illustrative purposes only. Those skilled in the art will recognize that some of the data in Figure 2 may be omitted or replaced with other data in the implementation of highlight identification and / or metadata extraction. Additionally or alternatively, data not shown in Figure 2 may be used in the implementation of highlight identification and / or metadata extraction.

[0116] Card image Next, referring to Figure 3, a screenshot diagram of an example of a video frame 300 from a video stream with information embedded in the form of card images, as frequently seen in television programs of sporting events. Figure 3 depicts a card image 310 in the lower right of video frame 300, and a second card image 320 extending along the bottom of video frame 300. Card images 310 and 320 may contain embedded information such as the game phase, the current clock, and the current score.

[0117] In at least one embodiment, information within card images 310, 320 is localized and processed for automatic recognition and interpretation of embedded text within card images 310, 320. The interpreted text may then be assembled into text metadata describing the status of the sports match at a specific point in time within the timeline of the sports event.

[0118] In particular, card image 310 may relate to the sporting event currently being shown, while a second card image 320 may contain information about a different sporting event. In some embodiments, only card images containing information deemed to be related to the sporting event currently being played are processed for metadata generation. Therefore, without limiting the scope, the following exemplary description assumes that only card image 310 is processed. However, in alternative embodiments, it may be desirable to process multiple card images within a given video frame 300, including card images related to other sporting events.

[0119] As shown in Figure 3, the card image 310 can provide several different types of metadata 224, including team name 330, score 340, leading team performance 350, current game stage 360, game clock 370, play status 380, and / or other information 390. Each of these may be extracted from within the card image 310 and interpreted to provide highlights 220 including video frames 300, more specifically, metadata 224 corresponding to the video frame 300 on which the card image 310 is displayed.

[0120] Metadata Extraction Figure 4 is a flowchart illustrating a method 400, according to one embodiment, which is performed by an application running on, for example, a client device 106 and / or an analysis server 116, where the method 400 receives a video stream 222 and performs on-the-fly processing of video frames 300 to extract metadata from card images such as card image 310. System 100 in Figure 1A is referred to as the system that performs method 400 and subsequent systems. However, alternative systems, including but not limited to system 160 in Figure 1B, system 180 in Figure 1C, and / or system 190 in Figure 1D, can be used instead of system 100 in Figure 1A.

[0121] Method 400 in Figure 4 illustrates the process outlined above in more detail. A video stream, such as video stream 222 corresponding to a previously identified highlight 220, may be received and decoded. In step 410, one or more video frames 300 of video stream 222 may be received, resized to a standard size, and decoded. In step 420, the video frames 300 may be processed to detect one or more card images from the video frames 300, such as the card images 310 in Figure 3, and extract them if applicable. If no valid card image 310 is found in the video frames 300 according to query 430, Method 400 may return to step 410 to decode and analyze a different video frame 300.

[0122] If a valid card image 310 is found, step 440 may further process the video frame 300 to locate, extract, and process the detected card image 310, and extract and process any text boxes and / or strings of characters embedded in the card image 310. If no valid string 200 is found in the card image 310 according to query 450, method 400 may return to step 410 to process a new video frame 300.

[0123] If a valid string 200 is found in the card image 310, method 400 can proceed to step 460, where the extracted string(s) 200 is recognized and interpreted, and corresponding metadata 224 is generated based on the interpretation of the information from the card image 310. In various embodiments, the available choices for text interpretation are based on determining the type of card image of the card image 310 detected in the video frame 300, and / or prior knowledge of the detected fields present in a particular type of card image applicable to the card image 310 detected in the video frame 300.

[0124] As previously shown, the detection, localization, and interpretation of embedded text in card images present in television program content may be performed entirely locally on the television, STB, or mobile device. Alternatively, it may be performed remotely on a remote video server with broadcast video capture and streaming capabilities. Or, any combination of local and remote processing can be used.

[0125] String processing of information cards: Location identification and extraction An "extreme region" (ER) is an image region where the outer boundary pixels have values ​​strictly higher than the region itself (e.g., Neumann, J. Matas, "Real-Time Scene Text Localization and Recognition", 5th IEEE Conference on Computer Vision and (Pattern Recognition, Providence, RI, June 2012). One of the well-known methods used for ER detection in images is the use of a so-called maximum stable ER detector or MSER detector. Additional detection methods allow for the examination of a wider range of ERs while keeping computational complexity relatively low. When a wider range of ERs is to be examined, a sequential classifier based on specific features related to character regions can be introduced. This classifier can be pre-trained to generate probabilities of character presence, resulting in multiple assumed detected boundaries for the character (i.e., character boundaries 204). In the first stage of ER classification, the probability of character presence is estimated, and in the second stage, the ER with the highest probability locally is selected. Classification can be further improved by using several more computationally intensive features. Furthermore, in at least one embodiment, iterative and exhaustive searches are applied to detect character combinations and group ERs into words. Such methods can also include region edges as ER considerations to improve character detection. The final result is the ER selected with the highest probability representing character boundary 204.

[0126] Since the character detector described above generates multiple regions for the same character, the next step is to perform ambiguity resolution on the detected regions. In at least one embodiment, this ambiguity resolution includes performing multiple comparisons of the detected character boundaries 204 and then purging any character boundaries 204 that may appear to be in the form of character bounding boxes that are too close to each other. As a result, only one character bounding box is accepted within a given perimeter, thus enabling the correct formation of the string 200 representing the appropriate text field of the card image 310.

[0127] Figure 5 is a flowchart illustrating method 500 for performing the process outlined above in more detail. Video frames 300 are selected for processing, or an option is selected to process each video frame 300 sequentially. In step 510, if a card image 310 is detected within the video frame 300, it is extracted and resized to a standardized size. Next, in step 520, the resized card image is preprocessed with a series of filters, including, for example, contrast enhancement, bilateral and central filtering for noise reduction, gamma correction, and / or lighting compensation.

[0128] In step 530, an ER filter with a two-stage classifier is created, and in step 540, this cascade classifier is applied to each image channel of the card image 310. Character groups are detected, and groups of one or more word boxes are extracted for further processing. In step 550, the strings 200 with individual character boundaries 204 are analyzed for character boundary deambiguation. Finally, a clean string 200 is generated, in which only one character is accepted within each of the character positions 206.

[0129] Figure 6 is a flowchart illustrating method 600 for further processing to verify character boundaries 204. Method 600 can begin in step 610 with the extraction of the string 200, the removal of duplicate characters, and the final processing and acceptance of the string 200. As shown, each character in the deambiguated string can be further processed for character image verification.

[0130] Therefore, in step 620, in the low-intensity and high-intensity regions of each character image 202 (or processed character image 203), the ratio of pixel counts can be obtained for comparison with a predetermined contrast ratio between the low-intensity and high-intensity pixel counts. In step 620, for each character image 202 or processed character image 203, pixels at high-intensity and low-intensity levels are grouped and counted.

[0131] Next, in step 630, the ratio of these two counts is calculated and thresholded so that only character images 202 or processed character images 203 with a sufficiently high contrast ratio are retained. Then, in step 640, the bounding box coordinates of the verified character (i.e., position 206) are recorded and stored for further use in the interpretation of the string 200.

[0132] In alternative embodiments, the character bounding box validation described above may precede character boundary ambiguation, or the validation may be used in combination with character boundary ambiguation for final character validation.

[0133] Information card processing for query extraction and recognition In at least one embodiment, an automated process is performed that includes the following steps: Receive a digital video stream, such as a video stream 222 associated with a highlight 220. Analyze one or more video frames 300 of the digital video stream for the presence of a card image 310. Extract the card image 310. Locate the character boundaries 204 of the characters in the string 200 within the card image 310. Extract the text within the text boxes to create a character query column.

[0134] Figure 7 is a flowchart illustrating a method 700 for generating an information card query according to one embodiment. In step 710, the card image 310 is extracted from the decoded video frame 300. In step 720, the card image 310 is processed to identify and extract the string 200 as described above. In step 730, the character image 202 is extracted from the card image 310 to generate a normalized query image (e.g., query 230). In step 740, the query infrastructure is input with the normalized query character image (query feature vector 232).

[0135] In another embodiment, query prediction is generated by first projecting the query feature vector onto a pre-developed training set of orthogonal basis (e.g., orthogonal basis 234), and then applying the resulting projected query feature vector to a machine learning-trained classification model such as classification model 165. The predicted alphanumeric string may be generated according to a pre-established classification label, and this predicted alphanumeric string may be passed to an interpretation process for the best extraction of the text's meaning 238.

[0136] Figure 8 is a flowchart illustrating method 800, which includes processing steps for query recognition, leading to query alphanumeric string generation and query interpretation and understanding. In step 810, orthogonal basis vectors of orthogonal basis 234 are loaded across the training set vector space. In step 820, the normalized query may be projected onto the orthogonal basis 234. In step 830, a classification model 165, such as a pre-developed one, may be loaded. The classification model 165 may be applied to the projected query. Finally, in step 840, a predicted alphanumeric string may be generated, which is then used for interpretation and semantic extraction to generate the text's meaning 238.

[0137] Query interpretation and semantic extraction In at least one embodiment, one or more strings 200 present within the card image 310 are identified. Subsequent steps may include locating, sizing, and extracting each character image 202 within the identified strings 200. The detected and extracted character images 202 are transformed into a query feature vector 232 and projected onto a training set orthogonal basis 234. The projected query is then applied to a classification model 165 to produce a predicted sequence of alphanumeric characters.

[0138] In at least one embodiment, the predicted query alphanumeric characters are sent to an interpretation process that applies prior knowledge and positional understanding of the characters present in a number of card images 310. Next, meaning is derived for each predicted alphanumeric character, positioned within a specific string 200, and the extracted information is appended to metadata 224 stored in association with the highlight 220.

[0139] Figure 9 is a flowchart illustrating in more detail a method 900 for predictive query column interpretation according to one embodiment. Method 900 includes combining consideration of string length, character box position and horizontal distance, and alphanumeric readings for semantic extraction.

[0140] Method 900 begins with step 910, in which the character count for each processed query for string 200 is loaded along with the size 205 and position 206 of the characters within string 200. The video frame number and / or time associated with the extracted query 230 being processed may also be made available for reference related to absolute time. In step 920, the column length 209, character size 205, and / or character position 206 may be taken into consideration in the analysis.

[0141] Next, in step 930, the system 100 advances the string 200, which may be interpreted by applying knowledge of the field positions of the characters, as well as knowledge of the alphanumeric values ​​of the characters. In step 930, knowledge and understanding of a particular card image 310 may also be used to compensate for any potentially missing leading characters. Finally, in step 940, the derived meaning is recorded (e.g., text meaning 238), and the corresponding metadata 224 is formed, providing real-time information related to the current timeline associated with the current sports event television program and the processed embedded card image 310.

[0142] Generation of machine learning-based classification models, involving the application of query character recognition to embedded information cards. In at least one embodiment, the generation of a classification model is performed using a convolutional neural network. Generally, the neural network develops an information classification function using known (desired) classification results through a supervised learning process applied to a training set of character vectors. During the training process, the neural network's algorithmic structure adjusts its weights and biases to perform accurate classification. An example of a known architecture used to learn the internal weights and biases of a neural network during the training process is the backpropagation neural network architecture, or feedforward backpropagation neural network architecture. When a set of training data is presented to such a network, the backpropagation algorithm calculates the difference between the actual output and the desired output, and feeds back the error to correct the internal network weights and biases that are causing the error. In the classification / inference phase, the neural network structure is first loaded with pre-trained model parameters, weights, and biases, and then queries are fed forward through the network, resulting in the output of one or more identified labels representing predictions for the queries.

[0143] Another exemplary system for classification model generation employs multi-class SVMs. Such SVM classification systems differ fundamentally from equivalent approaches, such as neural network learning systems, which heavily rely on heuristics to construct various network architectures, and whose training process does not always end at a global minimum. In contrast, SVMs are mathematically very clearly defined and have a training process that consistently finds a global minimum. Furthermore, using SVMs provides a relatively simple and clear geometric interpretation of the training process and classification objectives, which enhances intuitive insight into the process of classification model generation. SVMs can be efficiently used for classifying linearly inseparable datasets and can be extended to multi-label classification tasks. SVMs for classifying linearly inseparable datasets are characterized by the choice of kernel function that helps project the dataset into a high-dimensional vector space where the original dataset becomes linearly separable. However, the choice of kernel function is crucial and involves some degree of heuristics and data dependency.

[0144] In at least one embodiment, character classification model generation is based on a training set of characters extracted from one or more exemplary card images 310 embedded in television content of a sporting event. Character boundaries 204 are detected and characters are extracted from a number of card images 310. Such character boundaries 204 include small character images 202 that can then be normalized to standard size and illumination to provide processed character images 203. Feature vectors (or query feature vectors 232) are formed for the character images 202 and / or processed character images 203, and these feature vectors are then associated with each particular character from the set of character images appearing in the embedded card images 310.

[0145] In a structural approach to character image feature generation, the character feature vector, or query feature vector 232, is associated with a set of n pixels extracted from preprocessed character images 202. These n pixels are formatted into n-dimensional vectors representing single points in the n-dimensional feature vector space of the training vector. The primary goal of feature selection is to construct a decision boundary in the feature space that correctly separates character images 202 of different classes. Thus, in at least one embodiment, the set of extracted character images 202 representing the training vector is further processed to increase the uniqueness and cross-metric distance of the training vector, as well as to reduce the overall dimensionality of the vector space of the training vector.

[0146] In accordance with the above considerations, in another embodiment, principal component analysis (PCA) is performed on the training vector set. Thus, the orthogonal basis vectors of orthogonal basis 234 are derived from the training set such that the orthogonal basis vectors span the training vector space. Furthermore, the dimensionality of the training vector space is reduced by selecting a limited number of orthogonal basis vectors so that only the most important orthogonal vectors associated with the largest set of singular values ​​(generated by singular value decomposition of the basis vector matrix) are retained. The selected training set basis vectors are saved for later use in generating a classification model using one or more of the algorithmic structures available for dataset classification, such as an SVM classifier or a CNN classifier.

[0147] In various embodiments, the systems and methods described herein provide techniques for extracting individual character images 202 from a string 200 embedded in a card image 310, and then using the character images 202 to generate query feature vectors 232. In the next processing step, these query feature vectors are projected onto an orthogonal basis 234 spanning a training vector space to generate projected queries. The projected queries are then applied to generate query predictions or predicted queries 236 as the output of a pre-trained classification model produced by an exemplary SVM (or CNN) classifier. These predicted queries 236 form a sequence of predicted characters, which is then interpreted to generate text meaning 238 and finally used to generate metadata 224 of a highlight 220 enhanced with real-time information read directly from the card image 310.

[0148] Figure 10 is a flowchart illustrating the classification model generation method 1000 in more detail. In at least one embodiment, the method 1000 begins in step 1010, where an exemplary training set of character images 202 is extracted from a number of exemplary card image types. The character images 202 are normalized to standard size and illumination to form processed character images 203. Feature vectors are derived to generate a labeled training set. In at least one embodiment, in step 1020, PCA analysis is performed on the training set by computing orthogonal basis vectors 234 across the training vector space. In step 1030, a subset of orthogonal training vectors is selected. The selected training set basis vectors may be saved for query processing in step 1040. In step 1050, a classification model 165 may be trained on the subset of orthogonal training vectors. The classification model and orthogonal basis vectors may be saved in step 1060 for the generation of future predicted queries 236.

[0149] Figure 11 is a flowchart illustrating the overall method 1100 for reading and interpreting text fields within a card image 310 and updating metadata 224 for the highlight 220 with real-time information within the frame. In step 1110, the fields to be processed are selected from the character boundaries 204 of characters present in the card image 310. In step 1120, groups of characters are extracted from the row fields, the text columns are recognized and interpreted as described above. Finally, in step 1130, the reading of the card image performed at the decoded video frame boundary is embedded in the metadata 224 generated for the highlight 220.

[0150] The system and method have been described in particular detail with respect to assumed embodiments. Those skilled in the art will understand that the system and method may be implemented in other embodiments. First, the specific naming of components, the use of capitalization of terms, attributes, data structures, and any other programming or structural aspects are neither essential nor important, and mechanisms and / or functions may differ in name, format, and protocol. Furthermore, the system may be implemented through a combination of hardware and software, or entirely within hardware elements, or entirely within software elements. Also, the specific division of functions among the various system components described herein are merely illustrative and not essential. Functions performed by a single system component may instead be performed by multiple components, and functions performed by multiple components may instead be performed by a single component.

[0151] Any reference in this specification to “one embodiment” or “embodiment” means that a particular feature, structure, or characteristic described in relation to an embodiment is included in at least one embodiment. The phrases “in one embodiment” or “in at least one embodiment” appearing in various places in this specification do not necessarily all refer to the same embodiment.

[0152] Various embodiments may include any number of systems and / or methods for carrying out the above-described techniques, either individually or in any combination. Another embodiment includes a non-temporary computer-readable storage medium and a computer program product including computer program code encoded on the medium, for causing a processor in a computing device or other electronic device to carry out the above-described techniques.

[0153] Some of the above sections are presented in terms of algorithms and symbolic representations of operations on data bits in the memory of a computing device. These descriptions and representations of algorithms are means used by those skilled in the art of data processing to most effectively communicate the nature of the work to others skilled in the art. Here, an algorithm is generally considered to be a self-consistent set of steps (instructions) leading to a desired result. A step is one that requires the physical manipulation of a physical quantity. These quantities, though not always, take the form of electrical, magnetic, or optical signals that can be manipulated in ways such as storage, transfer, combination, comparison, and other methods. For reasons of common usage, it may be convenient to refer to these signals as bits, values, elements, symbols, characters, terms, or numbers, etc. Furthermore, without loss of generality, it may also be convenient to refer to a particular arrangement of steps requiring the physical manipulation of a physical quantity as a module or code device.

[0154] However, it should be kept in mind that all these and similar terms are associated with appropriate physical quantities and are merely convenient labels applied to those quantities. Unless otherwise stated, as will be evident from the following descriptions, throughout this specification, any use of terms such as “processing,” “computing,” “calculating,” “display,” or “determining” will be understood to refer to the operation and processes of a computer system or similar electronic computing module and / or device, and to mean manipulating and transforming data that is represented as a physical (electronic) quantity within the memory or registers or other such storage, transmission device, or display device of the computer system.

[0155] Certain embodiments include process steps and instructions described herein in the form of algorithms. Process steps and instructions can be embodied in software, firmware, and / or hardware, and if embodied in software, they can be downloaded to reside on various platforms used by various operating systems, and it should be noted that they can be operated from various platforms.

[0156] This document also relates to apparatus for carrying out the operations described herein. Such apparatus may include general-purpose computing devices that can be specifically constructed for a required purpose or selectively activated or reconfigured by computer programs stored in the computing device. Such computer programs may be stored in computer-readable storage media such as floppy disks, optical disks, CD-ROMs, DVD-ROMs, magneto-optical disks, read-only memory (ROM), random access memory (RAM), EPROMs, EEPROMs, flash memory, solid-state drives, magnetic or optical cards, application-specific integrated circuits (ASICs), or any type of medium suitable for storing electronic instructions, each coupled to a computer system bus. The program and its associated data may also be hosted and executed remotely, for example, on a server. Furthermore, the computing devices referred to herein may include a single processor or may be architectures employing multiple processor designs to enhance computing power.

[0157] The algorithms and representations presented herein are not inherently related to any particular computing device, virtualization system, or other apparatus. Various general-purpose systems may also be used with programs following the teachings herein, or it may prove convenient to construct more specialized apparatuses to carry out the required method steps. The structures required for these various systems will become apparent from the descriptions provided herein. Furthermore, the systems and methods are not described with reference to any particular programming language. Various programming languages ​​may be used to carry out the teachings described herein, and it will be understood that any above references to specific languages ​​are provided for the purpose of enabling and disclosing best modes of use.

[0158] Accordingly, various embodiments include software, hardware, and / or other elements, or any combination or combination thereof, for controlling a computer system, computing device, or other electronic device. Such electronic devices may include, for example, input devices such as a processor, keyboard, mouse, touchpad, trackpad, joystick, trackball, microphone, and / or any combination thereof; output devices such as a screen and / or speaker; long-term storage devices such as memory, magnetic storage devices, and / or optical storage devices; and / or network connectivity. Such electronic devices may be portable or non-portable. Examples of electronic devices that can be used to implement the described systems and methods include desktop computers, laptop computers, televisions, smartphones, tablets, music players, audio devices, kiosks, set-top boxes, game systems, wearable devices, home electronic devices, and / or server computers. Electronic devices may use any operating system, such as, for example, Linux®, Microsoft Windows (available from Microsoft Corporation, Redmond, Washington), Mac OS X (available from Apple Inc., Cupertino, California), iOS (available from Apple Inc., Cupertino, California), Android (available from Google Inc., Mountain View, California), and / or any other operating system adapted for use on the device.

[0159] While a limited number of embodiments have been described herein, those skilled in the art who appreciate the merits of the above description will understand that other embodiments may be conceived. Furthermore, it should be noted that the language used herein has been chosen primarily for readability and educational purposes and may not have been chosen to depict or limit the subject matter. Accordingly, this disclosure is intended to illustrate, but not to limit, the scope.

Claims

1. A method for extracting metadata from a video stream, wherein the method is The processor receives at least one portion of the video stream, The processor identifies one or more card images embedded in one or more video frames of the portion of the video stream, The processor processes one or more card images to extract text, In the aforementioned processor, Based on the character images in the extracted text, multiple query feature vectors are generated, each corresponding to a single character image. Based on the aforementioned multiple query feature vectors, a string consisting of multiple predicted characters is generated, The text is interpreted and metadata is obtained by interpreting the string consisting of the multiple predicted characters based on the positions corresponding to the multiple predicted characters within the string, A method comprising storing the metadata in a data store in association with the portion of the video stream.

2. The method according to claim 1, further comprising storing the received portion of the video stream in the data store.

3. The aforementioned video stream includes television broadcasts of sporting events, The portion of the video stream includes highlights that are deemed to be of particular interest to one or more users, The method according to claim 1, wherein the metadata describes the highlighting.

4. The method according to claim 3, further comprising outputting the video stream to an output device simultaneously with at least one of identifying the one or more card images, processing the one or more card images, and interpreting the text.

5. In the output device, outputting the highlights, The output of the highlights and the output of the metadata are further included, The aforementioned metadata, Real-time information related to the aforementioned highlights, The method according to claim 3, comprising at least one of the following: the metadata obtained from the card image and the timeline of the card image.

6. Extracting the aforementioned text is Identifying one or more strings within the aforementioned one or more card images, The method according to claim 1, comprising recording the position and / or size of the character images of the card images having one or more card images corresponding to each character of the one or more strings.

7. Extracting the aforementioned text is By performing multiple comparisons of detected character boundaries, the ambiguity of character boundaries between characters in one or more strings is resolved, The method according to claim 6, further comprising purging any character boundaries that appear when they are too close to each other.

8. The method according to claim 6, further comprising performing image verification on characters of one or more strings by extracting the text, thereby establishing a contrast ratio between low-intensity pixel counts and high-intensity pixel counts.

9. Interpreting the aforementioned text, Projecting the aforementioned query feature vector onto the orthogonal basis of the training set, The method according to claim 1, comprising applying the projected query feature vector to a classification model to produce at least one predicted query.

10. Generating multiple training set feature vectors, The method according to claim 9, further comprising deriving the orthogonal basis of the training set using the training set feature vectors.

11. Generating multiple training set feature vectors, The method according to claim 9, further comprising generating the classification model using the training set feature vectors.

12. Interpreting the aforementioned text, The length of one or more strings in the aforementioned text, The character boundaries and / or character positions within the aforementioned text, The method according to claim 9, further comprising using at least two selections of character boundaries and / or the horizontal position of characters within the text.

13. The method according to claim 9, wherein storing the metadata in association with the portion of the video stream includes storing the video frame numbers of one or more video frames associated with the query.

14. Interpreting the aforementioned text, The field position of one or more characters in the aforementioned text, Verify the alphanumeric value of the aforementioned characters, The method according to claim 1, comprising interpreting one or more strings sequentially using the field positions and alphanumeric values.

15. Interpreting the aforementioned text, To obtain location information and other information relating to one or more card fields of each of the aforementioned card images, The method according to claim 14, further comprising using the positional information and other information to compensate for one or more potentially missing leading characters of the one or more strings.

16. A system for extracting metadata from a video stream, Memory configured to store processor-readable instructions, The system comprises a processor coupled to the memory and configured to execute instructions and perform an operation, wherein the operation is Receiving at least one portion of the aforementioned video stream, Identifying one or more card images embedded in one or more video frames of the portion of the video stream, The process of processing one or more card images and extracting text, The text is interpreted and metadata is obtained by: generating multiple query feature vectors corresponding to each character image based on the character images in the extracted text; generating a string consisting of multiple predicted characters based on the multiple query feature vectors; and interpreting the string consisting of multiple predicted characters based on the positions corresponding to the multiple predicted characters within the string. A system comprising storing the metadata in a data store in association with the portion of the video stream.

17. The system according to claim 16, further comprising storing the received portion of the video stream in the data store.

18. The aforementioned video stream includes television broadcasts of sporting events, The portion of the video stream includes highlights that are deemed to be of particular interest to one or more users, The system according to claim 16, wherein the metadata describes the highlight.

19. A non-temporary computer-readable medium configured to store processor-readable instructions, wherein the instructions are executed by the processor. Receiving at least one portion of the video stream, Identifying one or more card images embedded in one or more video frames of the portion of the video stream, The process of processing one or more card images and extracting text, The text is interpreted and metadata is obtained by: generating multiple query feature vectors corresponding to each character image based on the character images in the extracted text; generating a string consisting of multiple predicted characters based on the multiple query feature vectors; and interpreting the string consisting of multiple predicted characters based on the positions corresponding to the multiple predicted characters within the string. A non-temporary computer-readable medium on which an operation is performed, including storing the metadata in a datastore in association with the portion of the video stream.

20. The non-temporary computer-readable medium according to claim 19, further comprising storing the received portion of the video stream in the data store.

Images