Methods and systems for identifying event participants in images of events.

Abstract

A computer-based method is described for identifying unique instances of participants with a common emblem in images of an event. The method stores specified parameter values ​​for features of each participant's emblem, and the relationship between unique combinations of these specified parameter values ​​and the identity of each participant in the current event. Images of the current event are cropped, depicting at least a portion of a participant's emblem; and the participant's identity is determined by calculating a matching score between the detection parameter values ​​of each emblem feature of each participant in the event and the stored specified parameter values. The participant's identity can then be determined using the best overall score that matches the stored parameter values ​​of the emblem of each participant in the event.

Description

Technical Field

[0001] This disclosure relates to a computer-based method and system for identifying participants in one or more images of an event using a trained classifier. Background Technology

[0002] In live video broadcasts or digital streaming of sporting events, it has become common practice for event organizers, broadcasters, or streaming channels to overlay information (such as participant identification numbers and / or sequential positions) onto the image stream to enhance the viewer experience and engagement. For example, in horse races or cycling races—especially in live (or streamed) or streaming events—the names and identifiers of the horses in a horse race or the cyclists in a cycling race can be displayed on the image of the event.

[0003] Typically, this information is overlaid on the video / image stream that constitutes the event and transmitted to all viewers. Unfortunately, this information may not be easily understood by some viewers. Furthermore, some viewers may only be interested in specific participants, but the video / image stream presents information about all participants.

[0004] This information is typically provided by human observers who manually examine images, identify participants, and encode information onto the images for transmission. However, such methods are cumbersome, time-consuming (and therefore expensive), and error-prone. Furthermore, such identification processes require specially trained operators who have limited speed at which they can encode information into the image stream.

[0005] Alternative approaches in some events involve using geographic tracking methods (such as GPS, RFID, and radio transmitters) to track the spatial location of participants at any given time. Typically, such methods require equipment mounted on participants, making deployment over large geographic areas or among numerous participants relatively expensive; and equipment malfunctions can occur at critical moments in the event.

[0006] This invention addresses or at least improves some of the above-mentioned deficiencies of the prior art, or at least provides further options for members of the public. Summary of the Invention

[0007] In a first aspect, the present invention provides a computer-based method for identifying at least one participant among a plurality of participants in one or more images of an image stream of an event, wherein each of the participants has a unique instance of a common emblem associated with it. The method includes receiving specified parameter values ​​of one or more features or groups of features constituting the emblem of each participant in the current event; storing a relationship between a unique combination of the specified parameter values ​​characterizing the emblem or groups of features and the identity of each participant in the current event; after the current event begins, cropping one or more image regions from one or more images of the image stream of the current event, at least depicting a portion of the emblem of one participant in the event; and predicting the identity of a participant among the participants in the event from the one or more images of the image stream of the event through the following steps:

[0008] (i) Determine a matching score between the detection parameter value of each feature or feature group of the corresponding portion of the emblem of each participant in the event and a stored specified parameter value; wherein the determination is performed by using a trained image classifier for each of one or more image regions in one or more images of the event and for each feature or feature group of each portion of the emblem of the participant depicted therein; and

[0009] (ii) The identity of a participant is determined by the best corresponding overall score that matches the detection parameter values ​​of the features or feature groups depicted in one or more images with the stored specified parameter values ​​of the features or feature groups of the emblem of each participant in the event.

[0010] Advantageously, the features or feature groups of the public emblem include one or more of pattern, shape, and color. The trained image classifier can be selected from the group of neural network classifiers including Inception V4 neural networks, VGG16 neural networks, ResNet neural networks, or convolutional neural networks, which can provide performance improvements during the training and service phases. In the first arrangement of the first embodiment of this disclosure, the score of the detection parameter value can be directly output from the trained image classifier.

[0011] The score comparison between the detection parameter value of each participant in the current event (for at least some of the features or feature groups of each participant) and the stored specified parameter value can be determined using the Hungarian algorithm.

[0012] Alternatively, the scores for the detection parameter values ​​can be computed from the classification classifier using the output of a trained image classifier. The optimal score for the parameter values ​​can be determined by multiplying the scores determined for the parameter values ​​of all features for each part of the common emblem associated with each participant in the event.

[0013] In the second arrangement of the first embodiment of this disclosure, the trained image classifier may be selected from the group consisting of Inception V4 neural networks, VGG16 neural networks, ResNet neural networks, or convolutional neural networks, which can provide performance improvements during the training and service phases.

[0014] The trained image classifier can be trained using an image classifier configured as a Siamese neural network; wherein the loss function of the Siamese neural network used for the trained image classifier is a contrastive loss function; and the input to the Siamese neural network is either an anchor image and a tuple of images having parameter values ​​similar to those of the anchor images in the set, or an anchor image and a tuple of images having parameter values ​​dissimilar to those of the anchor images in the set.

[0015] Advantageously, the accuracy of the image classifier can be improved by inputting images similar to and dissimilar to the anchor image during the training phase. In an alternative arrangement to the second arrangement of the first embodiment of this disclosure, an image classifier configured as a Siamese neural network can be used to train the trained image classifier; wherein the loss function of the Siamese neural network used for the trained image classifier is a triple loss function; and the input to the Siamese neural network is a tuple of the anchor image, images having parameter values ​​similar to those of the anchor images in the set, and images having parameter values ​​dissimilar to those of the anchor images in the set.

[0016] The score for comparing the detection parameter value of each participant in the current event (for at least some of the features or feature groups of each participant) with the stored specified parameter value can be determined based on the distance between the vector representation of the detection parameter value of the feature or feature group of each participant in the current event and the corresponding vector representation of the stored specified parameter value.

[0017] In a second embodiment of this disclosure, advantageously, model A is trained to map specified parameter values ​​to spatially distant locations, thereby improving classification accuracy. The specified parameter values ​​can be a vector representation of parameter values ​​for a feature or feature group of a part of a badge associated with a participant, generated through the following steps:

[0018] (i) Select one or more second trained classifiers for the feature or feature group; wherein each second trained classifier is trained to map similar vectors to be spatially relatively close in the output space of the second trained classifier, and to map dissimilar vectors to be spatially relatively far apart in the output space of the second trained classifier; and

[0019] (ii) Provide a second trained classifier selected for the feature or feature group with corresponding encoded parameter values ​​for one or more portions of the common emblem of the participants of the event to generate the vector representation.

[0020] Furthermore, in a second embodiment of this disclosure, an image classifier configured as a Siamese neural network can be used to train the trained image classifier; wherein the loss function of the Siamese neural network used for the trained image classifier is a contrastive loss function; and the input to the Siamese neural network is a vector representation of specified parameter values ​​and a tuple of images having parameter values ​​similar to those specified parameter values ​​in the set, or a vector representation of specified parameter values ​​and a tuple of images having parameter values ​​dissimilar to those specified parameter values ​​in the set.

[0021] Advantageously, the accuracy of an image classifier can be improved by inputting images similar to and dissimilar to vector representations of specified parameter values ​​during training. In an alternative arrangement of a second embodiment of this disclosure, an image classifier configured as a Siamese neural network can be used to train the trained image classifier; wherein the loss function of the Siamese neural network used for the trained image classifier is a triple loss function; and the input to the Siamese neural network is a tuple of vector representations of specified parameter values, images having parameter values ​​similar to those in the set, and images having parameter values ​​dissimilar to those in the set.

[0022] The second trained classifier can be trained using a classifier configured as a Siamese neural network, wherein the input to the Siamese neural network of the classifier includes: an anchor vector encoding parameter values; a vector encoding parameter values ​​that is similar to the anchor vector, including the anchor vector with added noise encoding the parameter values ​​or a vector obtained from the same combination of parameter values ​​as the anchor vector; and a dissimilar vector whose parameter values ​​are not similar to the anchor vector, including vectors with different combinations of parameter values ​​compared to the anchor vector.

[0023] One or more of the parameter values ​​in the encoded parameter values ​​can be encoded using one-hot encoding. Alternatively, one or more of the parameter values ​​in the encoded parameter values ​​can be encoded using RGB values, which advantageously provides greater robustness of the classifier when classifying color features. The loss function for the Siamese neural network used for the classifier can be chosen from either a contrastive loss function or a triple loss function.

[0024] The second trained classifier (Model A) can be selected from the group consisting of: a custom neural network, or any one of the Inception V4 neural network, VGG16 neural network, ResNet neural network, and convolutional neural network. Custom neural networks have the advantage of providing higher efficiency in training without compromising accuracy.

[0025] The public emblem associated with a participant may include items of clothing worn by the participant and / or parts of accompanying entities. Characteristics of the participant's emblem and / or accompanying entities may include patterns, shapes, and colors. The public emblem may include any one or more of a hat, body, sleeves, shorts, and boots. Accompanying entities of the participant may include any one or more entities selected from the group consisting of: horses, vehicles, bicycles, and motorcycles.

[0026] In a second aspect, the present invention provides a computer-based method for training a system for identifying at least one of a plurality of participants in one or more images in an image stream of an event, wherein each of the participants has a unique instance of a common emblem associated with it. The method includes receiving a set of images or portions thereof depicting emblems of participants in an ended event; wherein the depicted emblems have detectable parameter values ​​for features or feature groups of common emblems of participants in the ended event; wherein the set of images or portions thereof is labeled according to the detectable parameter values ​​of the emblems of the participants depicted therein; and generating a subset of training images from the set of images or portions thereof having specified parameter values ​​for one or more features or feature groups of emblems associated with each participant in the current event; training at least one classifier on the subset of training images for each feature or feature group to score the images or portions thereof for the specified parameter values ​​of the feature or feature group, wherein the training includes providing a corresponding classifier for each feature or feature group with input including the corresponding subset of training images; and iteratively updating the classifier for each feature or feature group until each classifier reaches a predetermined scoring accuracy.

[0027] Advantageously, the collection of images of participants or portions thereof from a concluded event may not include images having all parameter values ​​of the characteristics of the participants' emblems. Alternatively, the collection of images of participants or portions thereof from a concluded event may include multiple images of the characteristics of the emblem portions, which contain all parameter values ​​of the characteristics of the participants' emblems.

[0028] At least one classifier may be an image classifier selected from the group of neural network classifiers including: Inception V4 neural network, VGG16 neural network, ResNet neural network, or convolutional neural network.

[0029] In this disclosure, the training may be the training of two or more classifiers configured as a Siamese neural network, which uses a contrastive loss function, a triple loss function, or a cross-entropy loss function. The input to the Siamese neural network of the classifier may include a tuple of anchor images and similar images, or a tuple of anchor images and dissimilar images. Alternatively, the input to the Siamese neural network of the classifier may include a tuple of anchor images, similar images, and dissimilar images. Furthermore, the input to the Siamese neural network of the classifier may include anchor vectors, similar images, and dissimilar images.

[0030] The set of images may be generated after performing one or more image preprocessing steps, which are selected from the group including resizing and augmenting the acquired images.

[0031] Anchor vectors can be generated through the following steps:

[0032] (i) Select one or more second trained classifiers for the feature or feature group; wherein each second trained classifier is trained to map similar vectors to be spatially relatively close in the output space of the second trained classifier, and to map dissimilar vectors to be spatially relatively far apart in the output space of the second trained image classifier; and

[0033] (ii) Provide a second trained classifier selected for the feature or feature group with corresponding encoded parameter values ​​for features or feature groups corresponding to one or more portions of the public emblem of the participants in the event to generate the anchor vector.

[0034] The second trained classifier can be trained using a classifier configured as a Siamese neural network, wherein the input to the Siamese neural network of the classifier includes: a parameter vector encoding parameter values; a vector encoding parameter values ​​that is similar to the parameter vector, including a parameter vector encoding parameter values ​​with added noise or a vector obtained from the same combination of parameter values ​​as the parameter vector; and a dissimilar vector whose parameter values ​​are not similar to the parameter vector, including vectors that have different combinations of parameter values ​​compared to the parameter vector.

[0035] Advantageously, one or more of the parameter values ​​in the parameter vector can be encoded using one-hot encoding. Furthermore, one or more of the parameter values ​​in the parameter vector can be encoded using RGB values.

[0036] In a third aspect, a computer-readable medium including instructions is provided, wherein, when executed by a processor, these instructions cause the processor to perform any of the steps of the methods described above.

[0037] In a fourth aspect, a system is provided for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, wherein each of the participants has a unique instance of a common emblem associated with it, and the system includes an image processing module configured to perform the identification steps described above. Further, the system provides a training module configured to perform the training steps described above. Attached Figure Description

[0038] The preferred embodiments of this disclosure will be explained in further detail below with reference to the accompanying drawings, in which: -

[0039] Figure 1 High-level schematic representations of various embodiments of the present disclosure are depicted.

[0040] Figure 2 A flowchart illustrating the steps performed by an exemplary definition module according to a first embodiment of the present invention is provided.

[0041] Figure 3A A flowchart depicts the steps performed by an exemplary training module in a first arrangement according to a first embodiment of the present invention, and a first stage of a second arrangement according to a first embodiment of the present invention.

[0042] Figure 3B A flowchart depicts further steps for training a classification classifier performed by an exemplary training module in a second arrangement according to a first embodiment of the present invention.

[0043] Figure 4 A flowchart depicts the steps performed during the service or deployment phase of a first embodiment of the present invention.

[0044] Figure 5 The second arrangement depicting the first embodiment of this disclosure Figure 1 A more detailed example of the block diagram.

[0045] Figure 6A A flowchart depicts the steps performed during an exemplary training process according to a second embodiment of the present invention.

[0046] Figure 6B A block diagram depicts a module in an exemplary training process according to a second embodiment of the present invention.

[0047] Figure 7A flowchart depicts the steps performed during an exemplary training process of Model A according to a second embodiment of the present invention.

[0048] Figure 8 A flowchart depicts the steps performed during an exemplary training process of a B-model according to a second embodiment of the present invention.

[0049] Figure 9 A flowchart depicts the steps performed during an exemplary configuration process of a participant's A model according to a second embodiment of the present invention.

[0050] Figure 10 A flowchart depicts the steps performed during the service process of a second embodiment of the present invention.

[0051] Figure 11 A more detailed block diagram depicts the configuration process and service or deployment phase in a second embodiment of this disclosure.

[0052] Figure 12A Examples of jockey hats are depicted; jockey hats are part of the insignia associated with jockeys in horse racing.

[0053] Figure 12B Examples depicting the jockey's body section, which is part of the emblem associated with jockeys in horse racing.

[0054] Figure 12C Examples depicting jockey sleeves are shown; jockey sleeves are part of the insignia associated with jockeys in horse racing.

[0055] Figure 13 Some exemplary parameter values ​​describe the color characteristics.

[0056] Figure 14 Visual and parametric representations of exemplary parameter values ​​depicting characteristics of a portion of the emblem defined for a participant in an event (in this example, a jockey in a horse race).

[0057] Figure 15 An exemplary scoring table that describes the characteristics or groups of characteristics of the participants in an event.

[0058] Figure 16 An exemplary schematic representation of a computer system in which the steps of this disclosure can be performed is depicted. Detailed Implementation

[0059] This disclosure will now be described more fully with reference to the accompanying drawings.

[0060] In all respects, and generally speaking, this disclosure teaches a method and system for identifying the identities of participants in one or more images in an image stream of an event using a computer-based machine learning model.

[0061] In the embodiments disclosed herein, the process generally includes three main modules: a definition module 10, a training module 20, and a service module 30, which operate together during the training phase, as follows: Figure 1 What is depicted. Figure 5 A second arrangement, similar to and more detailed than the first embodiment, is described in the text.

[0062] Participants in an event suitable for image analysis performed by the methods and systems of this disclosure wear unique badges, which include various parts of the participant's body or otherwise associated with the participant. It should be understood that the portion of the badge associated with each participant can also be located on or otherwise associated with the participant in the event (e.g., horse racing, bicycle racing, motorcycle racing, running, or football) without limitation. The computer-implemented methods and systems of this disclosure can even be used to identify vehicles (e.g., racing cars) where participants can drive cars with unique, specific panels or portions thereof, which similarly form a "badge" associated with that participant.

[0063] Various unique features or groups of features (such as colors or patterns) can have unique parameter values ​​that can be used to identify participants using the methods and systems disclosed herein. Reference Figures 12A to 12C You can see non-limiting examples of subsets of parameters of various parts of a emblem that are associated with and pre-specified in the exemplary event type (in this case, a jockey in a horse race).

[0064] As further described herein, definition module 10 receives input from a user specifying which parts of the emblem can be associated with participants in a particular type of event. As will be further described, different event types require different trained image processing models trained on different source images; however, the general principles of the method remain consistent.

[0065] An insignia associated with a participant in a completed event comprises one or more parts. In some examples, such as a jockey in a horse race, the insignia can be defined in the definition module as including the jockey's attire, such as a hat, bodysuit, and sleeves.

[0066] Similarly, participants in a racing competition may possess certain plates (such as vehicle windshields, hoods, and wheel arches) or different car types, which can be interpreted as emblems associated with participants in that type of racing competition. Importantly, the emblem scheme for a specific event is fixed during the definition phase and is consistent in the model training module during the training phase and the service phase, as discussed further in this paper.

[0067] In definition module 10, a logo scheme for the various parts of the logos associated with participants in a specific type of event is fixed for that event type. A set 12 of images of participants in previous events of the same type with associated logos consistent with the defined logo scheme can be obtained from a suitable available image library. In other words, the logo scheme defines features and / or groups of features for the purpose of establishing a set of images of events for the training phase.

[0068] According to methods known in the art, images of participants and associated badges from a concluded event at different angles, distances, etc., can be obtained from still images or image streams. This image set 12 is further processed and labeled; and then used in training module 20 to train various classifiers 22, as further described herein. Badges of participants in concluded events used in training are referred to as "seen" badges.

[0069] Before a specific event begins, the definition module 10 stores the parameter values ​​of the characteristics / characteristic groups of emblems associated with the participants in the upcoming event in a data storage device or database.

[0070] Advantageously, also prior to the start of a particular event, a subset of images from the total set of participant images can be used to train a known feature / feature group classifier, as will be described further, for the parameters necessary to identify the features / feature groups of emblems associated with the participants in that event.

[0071] Therefore, the classifier is trained during the training phase before the event occurs; and can then be used by the service module 30 during the service or deployment phase to process images 32 of unknown participants, which can be captured from the image stream of the event, preferably during the event or with minimal time delay.

[0072] A trained classifier can generate predicted scores 34 for the features or feature groups of the participants depicted in the captured image 32.

[0073] Then, the matching module 40 can be used to interpret the scores and determine the identities of the participants 42 in the images 32.

[0074] The participants' identities can then be displayed on the screen using text labels, etc. It should be understood that these labels can identify participants chosen by viewers for special attention, as well as those who may occupy a specific sequential position in the event (e.g., first / second / third), thereby promoting viewer interest and engagement. Once identified, various other information about the participants (e.g., history, pictures, previous records, etc.) can also be displayed without limitation.

[0075] In the first embodiment, the steps of the training and service phases will now be described in more detail.

[0076] Specifically, Figure 2 A flowchart 200 depicts the various steps performed in the definition module 10 according to a first embodiment of the present invention.

[0077] In step 202, images of participants from past events (such as competitions or contests) are collected. These events must be of the same type as the upcoming event; and regarding these events, it is expected that the computer-based learning model will be used to identify the participants in the images.

[0078] The corpus of images of participant emblems in the set should include images of all distinguishing features or feature groups (such as patterns, shapes, and colors) of the emblem portion. The features and their parameter values ​​represent all detectable design elements available in the emblem scheme for the event.

[0079] Advantageously, the corpus of images of participant badges seen does not need to include badge images of all combinations of features of the badge scheme and feature / feature group parameters, and therefore, the number of participant badge images that need to be collected for machine learning training is greatly reduced.

[0080] The number of images required will depend on the training method and the design and type of participant badges.

[0081] In an exemplary embodiment for identifying participants in a horse race (in which case a logo scheme for each part of the jockey’s silk garment is used), approximately 2,000 to 4,000 unique images of the participant’s logo are collected.

[0082] In step 204, depending on the type of emblem scheme, one or more features or feature groups suitable for identifying the participants can be determined based on the emblem scheme specified for that event type.

[0083] In an example of identifying a jockey in a horse race, features that can be used for identification purposes include hat pattern, hat color, body pattern, body color, sleeve pattern, and sleeve color. In an exemplary embodiment for identifying a cyclist in a bicycle race, identifiable features or groups of features that can be used in a logo scheme include a helmet shape with a helmet shape color, a helmet pattern with a helmet pattern color, a body pattern with a body pattern color, and a shorts pattern with a shorts pattern color.

[0084] Figures 12A to 12C Examples of parameter values ​​for the jockey's insignia features—hat pattern, body pattern, and sleeve pattern—are shown. Examples of parameter values ​​for the color features are also provided. Figure 13 The symbol describes and represents one or more colors on a feature, including but not limited to black, blue, brown, gold, green, orange, pink, red, rose, royal blue, silver, turquoise, purple, white, and yellow. It should be understood that any number of colors can be used as parameter values, as long as their differences are detectable and distinguishable. Color features are parameterized; if the symbol scheme uses RGB values ​​or similar values ​​to represent colors, then the RGB values ​​or similar values ​​need to be parameterized as color labels. For example, the RGB value #FF0000 would be parameterized as "red". Furthermore, multiple colors can be represented relative to any feature, and these colors can be classified as primary, secondary, and tertiary colors, etc., according to their prominence on the feature, without limitation.

[0085] For example, such as Figure 12A As shown, for the hat pattern feature, the parameter values ​​for the feature include, but are not limited to, herringbone, ring, ring-2 colors, ring tiling, no pattern, star, stripe, stripe-3 colors, and stripe tiling. Other hat patterns can be used in the emblem scheme. Similarly, for Figure 12B The depicted main pattern is a characteristic whose parameter values ​​include, but are not limited to, grids, rhombuses, discs, two-color rings, multi-rings (normal), large dots, no pattern, dots, stripes, two-color stripes, three rhombuses, V-shapes, etc. In Figure 12C In the design, the sleeve pattern can be herringbone, multiple herringbone patterns, diamond, ring (2 colors), rings laid flat, plain, polka dots, stripes, etc.

[0086] Similarly, in the example of identifying a motorcycle in a motorcycle incident, the identifiable features of the emblem scheme may include the front pattern of the motorcycle, the body pattern of the motorcycle, the rear pattern of the motorcycle, and features of the motorcyclist's emblem.

[0087] Similarly, the characteristic colors associated with identifying riders and motorcycles are the colors on detectable features of parts of the participant or accompanying entity, including but not limited to, such as... Figure 13The colors depicted are black, blue, brown, gold, green, orange, pink, red, rose, royal blue, silver, turquoise, purple, white, and yellow, or any detectable and distinguishable color.

[0088] In step 206, depending on the source of the images of the participants' badges seen (e.g., video footage or still images of past competitions or events), the source images may need to be processed to obtain images for the image set. Advantageously, this preprocessing may include steps 206(a) to (c) as described herein.

[0089] In step 206(a), each participant depicted in the image can be detected by a first object detector (object detector model 1), which is configured to crop the image of each participant based on the coordinates returned by the first object detector using machine vision.

[0090] Advantageously, the first object detector is an AI model that can be specifically trained to detect certain objects, and this object detector can be trained based on open-source or existing object detector models (such as SSD and YOLO). Alternatively, labeled images of detectable objects can be used to train the new object detector. The first object detector can be trained to detect regions of an image that contain participants or one or more parts of those participants in the image. For example, in the case of jockeys and cyclists, image regions containing parts of the competitor (e.g., the participant's head, body, arms, and lower body) can be cropped by the object detector, and these parts include unique signature features. In different cases, the first object detector can be trained to crop other parts of the participant that contain unique features / feature groups.

[0091] Next, in step 206(b), for the obtained image of the participant, for each portion of the participant's emblem that includes at least one feature, one or more further object detectors may be used. These further object detectors may be trained to detect one or more specific portions of the participant's emblem and to crop an image region or a portion thereof containing that one or more portions of the participant's emblem to obtain an image portion.

[0092] Advantageously, the object detectors used to detect parts of the participant's emblem can be specialized object detectors for detecting parts of the emblem (such as hats, body parts, sleeves, and shorts), and are referred to as object detector p1, object detector p2, and so on, up to object detector pN, for parts p1, p2, ..., pN. Similar to the first object detector, object detectors p1 to pN can be trained based on an existing model or using labeled images of the objects.

[0093] Alternatively, in step 206(c), the image of the emblem and its portion containing at least one feature can be further processed by augmentation (for expanding the training dataset) and / or resizing.

[0094] In step 208, based on the emblem scheme for that event type, parameter values ​​uniquely associated with each feature or group of features are defined for that event type. For example, suppose the jockey's emblem scheme is defined as depending on one feature (for simplicity), the feature hat pattern. In this case, all jockey participants in that event type will have similar but uniquely distinguishable patterns on their hats. The possible patterns for this feature will be defined with various parameter values, such as... Figure 12B Those depicted in the text. In step 210, the corpus of participant images obtained in step 206(a) and the corpus of images of each part of the emblem associated with the participant obtained in step 206(b) are classified into categories and labeled using various unique parameter values ​​of the characteristics of the emblem scheme.

[0095] In a further example, in a logo scheme that uses a portion of an image of a participant's emblem featuring a striped body pattern and a yellow body color, the image can be labeled {body pattern: stripes, body color: yellow}. This labeled image can be applied independently to the feature body pattern and body color; while the feature group with a colored body pattern can be labeled {colored body pattern: striped yellow}.

[0096] Figure 3A A flowchart 300 depicts the training period of a first arrangement of a first embodiment of the present disclosure and the first stage of a second arrangement of the first embodiment. In this arrangement of the first embodiment, an image classifier is trained directly on the image.

[0097] In step 302, a dataset is loaded for processing. This dataset includes images of participants from completed events of the same type as upcoming events and with the same logo scheme associated with the participants. For example, if analyzing upcoming horse racing events, the logos of participants with the same logo scheme from completed horse racing events would be relevant.

[0098] Advantageously, if the parameter values ​​of the emblem features of all participants in the upcoming event are known, the training time can be reduced, meaning that only step 210 (from Figure 2 The training is performed on a subset of the dataset saved in the process described in the text.

[0099] In the example mentioned earlier, for an upcoming horse race, the insignia parameter values ​​for all jockeys are known in advance. For that particular race, no jockey wears an insignia with a diamond pattern for the body pattern feature, and it is not necessary to include image categories labeled with the parameter value {body pattern: diamond} or equivalent values ​​in the training set of the body pattern classifier.

[0100] Similarly, suppose in the example, the jockey's badge scheme in horse racing defines 20 patterns or categories for specific features of the hat pattern. Each pattern or category of the hat pattern can be represented by parameter values, such as from... Figure 12A Choose from those depicted in the text.

[0101] If, in an upcoming race, the jockey will only have 5 hat patterns or categories, then the classifier can simply be trained on images representing the parameter values ​​of those 5 categories, using images that independently represent the features / feature groups.

[0102] However, for each feature, the superset of the image dataset should at least contain image categories representing all parameter values ​​of the features in the emblem scheme that will appear in that type of event. (That is, the superset of images should preferably contain at least diamond patterns, but the training set does not necessarily have to contain large diamond patterns.) Therefore, a subset of the image dataset used to train a classifier for jockeys in a particular race can be constructed from images of participant jockey emblems that do not have the exact same combination of feature parameter values ​​as participants in an upcoming race.

[0103] In other words, when training a classifier for images that will be used to process upcoming events, it is not necessary to include images of logos of many past or synthetic participants that will be unique combinations of parameter values ​​of features on the logos of participants who will appear in the upcoming event.

[0104] In step 304, for the known parameters of the participants in the upcoming event, an image classification classifier is trained for each feature or feature group according to the emblem scheme of the event.

[0105] In step 306, an image classification classifier utilizing fully connected dense layers can be defined, which includes outputs for each parameter value of the feature. The image classification classifier can be a custom convolutional neural network or one of existing neural networks used for image classification (such as Inception V4, VGG16, and ResNet).

[0106] In step 308, the dataset is split into a training dataset and a test dataset. A common approach is to randomly split the entire dataset into a 9:1 or 8:2 ratio for training and testing.

[0107] In step 310, batch size, epochs, loss function (such as cross-entropy loss function or appropriate loss function) and other hyperparameters are configured for the image classification classifier defined in step 304.

[0108] In step 312, the image classifier is trained by inputting training images that are relevant to the definition of features or feature groups.

[0109] Therefore, in the first arrangement of the first embodiment, a trained image classifier 22 is obtained directly for the feature or feature group, such as... Figure 1 As depicted. The training process is repeated for each feature or feature group in steps 304 to 314 to obtain a set of trained image classifiers in step 316 for identifying participants based on parameter values ​​of the feature / feature group of the emblem portion that is typically associated with all participants in this event type.

[0110] As will be further described herein, particularly with reference to Figure 5 In a further arrangement of the first embodiment, the Siamese network of the image classifier can be used for training in step 304. In step 306, the Siamese network of the image classifier is configured, and the output of the image classifier is coupled through a fully connected dense layer, which in turn is coupled to the loss function. Steps 308 and 310 are the same as those discussed above for the first arrangement of the first embodiment.

[0111] In this further arrangement of the first embodiment, at step 312, the image classifier can be trained by inputting anchor images and similar image pairs with the same or similar parameter values, as well as anchor images and dissimilar image pairs with different or dissimilar parameter values.

[0112] Alternatively, such as Figure 5 The training of the Siamese network for the image classifier 522 at step 312, as described, can be performed by inputting a tuple comprising anchor images, similar images, and dissimilar images. The weights and biases of the image classifier are updated via backpropagation; and the model converges after several training epochs.

[0113] Figure 3BA flowchart 350 depicts the second stage of the training period of the second arrangement of the first embodiment, following the training of the image classifier and the acquisition of the image classifier from the image classifier's Siamese neural network. The classification classifier is trained to convert the vector output of the image classifier into a classification output of parameter values. Those skilled in the art will understand that any classifier suitable for classifying vector inputs can be used as the classification classifier. In step 352, a dataset consistent with the dataset used in step 302 is loaded for processing. This dataset includes images of participants in completed events of the same type as the upcoming events and has the same symbol scheme associated with the participants.

[0114] Advantageously, if the parameter values ​​of the emblem features of all participants in the upcoming event are known, training time can be reduced and the accuracy of the classifier can be improved, and in such a case, only step 210 (from Figure 2 The training is performed on a subset of the dataset saved in the process described in the text.

[0115] In step 354, one or more classification classifiers are defined for each feature or feature group of the emblem scheme based on the upcoming event, corresponding to each of the one or more image classification classifiers in step 304. These one or more classification classifiers are selected from any one of support vector machine classifiers, Bayesian classifiers, neural networks, or similar classifiers capable of classifying vectors in a multidimensional space. As those skilled in the art will understand, many classification techniques and algorithms are applicable, and depending on the parameter values, the classification classifiers can employ arrangements and modifications known in the art to perform multi-class classification.

[0116] In step 356, for each image classifier and its corresponding one or more classification classifiers, the output of the image classifier is configured as the input of the one or more classification classifiers.

[0117] In step 358, the dataset is split into a training dataset and a test dataset. A common approach is to randomly split the entire dataset into a 9:1 or 8:2 ratio for training and testing.

[0118] In step 360, the parameters of the classification classifier (such as the kernel of a support vector machine classifier) ​​and hyperparameters are defined in step 354. In the example of jockey badge classification, the radial basis function kernel provides a good fit to the data. It should be understood that those skilled in the art can choose a suitable kernel known in the art.

[0119] In step 362, training of one or more classifiers for each feature or feature group is performed for the part of the emblem scheme for the event by inputting the training images and parameter values ​​of the images defined by the relevant parameters into an image classification classifier for the feature or feature group as configured in step 356.

[0120] In step 364, after iterative training for each feature or feature group, a set of support vector machines corresponding to the image classifier is obtained.

[0121] In the second arrangement of the first embodiment, the trained image classifier is either image classifier 522a or 522b within the Siamese network, and is used to train the corresponding classification classifier 560, such as... Figure 5 As depicted, the first phase of the training process is repeated for each feature or feature group in steps 304 to 314 to obtain a set of trained image classifiers in step 316, and the second phase of the training process is repeated for each identical feature or feature group in steps 354 to 362 to obtain a set of classification classifiers 560 in step 364, which are used in conjunction with the corresponding trained image classifiers 520.

[0122] In the exemplary application of the methods and systems disclosed herein for identifying horse jockeys, separate image classification classifiers can be trained for features such as hat pattern, hat color, body pattern, body color, sleeve pattern, and sleeve color.

[0123] Alternatively, in the same application, but in an alternative arrangement, a separate image classification classifier can be trained to identify feature groups from images of the jockey, including colored hat patterns, colored body patterns, and colored sleeve patterns. The decision regarding which type of image classifier (feature or feature group classifier) ​​to use can be based on the combination of features worn by the participants in the upcoming race.

[0124] Figure 4 A flowchart 400 illustrating the service or deployment phase of a first embodiment of the present invention is shown.

[0125] During the service period, the corresponding trained image classifier is used to identify participants based on the unique recognition parameter values ​​of the features / feature groups of the emblem portion detected in the image or a part of the image.

[0126] In embodiments of this disclosure, a service procedure is performed on one or more images. It should be understood that the one or more images being analyzed can be extracted from a live stream, a pre-recorded video with frames arranged in chronological order, or another type of image stream.

[0127] For ease of explanation, assume that in step 402 a single image of the event of interest is extracted or accessed from the image stream. (Once multiple images from the video / image stream have been acquired and processed, the same steps will of course be repeated for these images).

[0128] In step 404, the object detector (object detector model 1) configured for participant detection is applied to the source image to locate the coordinates (bounding boxes) of each participant in the source image.

[0129] In step 406, based on the coordinates of each participant in the source image, the image of the located participant is cropped and stored in a database on a memory (such as a data storage device) or any form of computer-readable storage medium (such as a disk drive, memory chip, etc.) in any suitable form.

[0130] In step 408, for each image of the participant obtained from step 406, one or more object detectors (object detector p1 to object detector pN) for partial detection are applied to the image of the participant to locate the coordinates of each part of the participant corresponding to a feature or feature group.

[0131] In step 410, for each image or portion of an image of the located and cropped participant, a trained image classifier is applied to predict a score s for each parameter value of the feature or feature group associated with that portion and feature or feature group. f c In the second arrangement of the first embodiment, the trained image classifier is combined with the corresponding classification classifier to predict the score s. f c .

[0132] In the embodiments disclosed herein, the score s of the parameter values ​​of the feature or feature group f c It is an integer ranging from 0 to 1, and the score s for all parameter values. f c The sum (i.e., Σs) f c ) equals 1, where c is the parameter value of feature f.

[0133] In step 412, based on the score s for each unique feature f c Where f = {f1, f2, ..., fN}, the overall score s determines the parameter values ​​of the emblems of known participants, including their features. u .

[0134] For example, in the example of identifying a jockey, if there are 6 features, namely hat pattern, hat color, body pattern, body color, sleeve pattern, and sleeve color, then the score is obtained by evaluating the parameter values ​​of all features specific to the participant's insignia—hat pattern, hat color, body pattern, body color, sleeve pattern, and sleeve color. f c To calculate the score s of the participant's badge u .

[0135] In an implementation of this disclosure, in step 412, the overall score s for each badge of the participant is calculated. u This can be obtained by multiplying the individual scores of all the features of the emblem, i.e., s u =s f1 c1 s f2 c2 s f3 c3 .. s fN cN , where f1, f2, ..., fN are features, and c1, c2, ..., cN are parameter values ​​{a, b, c...} of features specific to a part of a logo or a common logo.

[0136] In another embodiment of the first embodiment of this disclosure, the score s can be... f c Apply logarithmic transformation to obtain log(s) f c This can be summed to calculate the logarithmic score, thus avoiding the multiplication of a large number of fractions. Other variations of probability addition or multiplication are also possible to avoid the overall probability being zero.

[0137] Furthermore, in alternative implementations, other algorithms (such as probabilistic or statistical methods, Hungarian algorithms, or neural networks) can be used to evaluate the scores. f c In order to obtain the overall score s in step 412 u However, this does not deviate from the scope of this disclosure.

[0138] In step 414, based on the overall score of all badge portions associated with the participants in the event, the best match for each participant in the event among all known badges in that event is determined (based on the best match score s). u Based on a badge that matches each participant, the participant's identity and other information (such as participant ID, history, last winner, etc.) can be identified and displayed to viewers of the image / image stream.

[0139] As an explanation, Figure 15 A matching score table is shown, with three features and three parameter values ​​for each feature. If participant 1 has the feature and parameter value {f1 = a...} f1 f2=a f2 f3 = c f3}, then having a score {a f1 =0.8,a f2 =0.8,c f3 The participant with the highest score among all participants is identified as participant 1, whose logo feature combination 1 is 0.8.

[0140] Advantageously, depending on the event's emblem scheme, participant identification is also possible without evaluating scores for all features of the emblem portion associated with the participant.

[0141] If only one participant among all participants in the event is wearing a black hat, then the participant identified by the black hat can be identified as that participant without further evaluation of parameter values ​​for other characteristics of the common / public parts of the emblem.

[0142] A further advantage is that it can still be identified even when some features of the participant are partially obscured or blurred in the image.

[0143] refer to Figure 1 In the arrangement according to the first embodiment, for each feature or feature group, multiple images 12 grouped under a category managed by the definition module 10 according to a parameter definition based on a unique combination of parameter values ​​are fed into the training module 20, in which one or more image classification classifiers are trained according to the features or feature groups defined by the parameter definition.

[0144] Figure 5 A schematic block diagram illustrating a further arrangement of a first embodiment of the invention is shown, wherein the image classifier is trained using a Siamese neural network as described above.

[0145] Similarly, in this arrangement, the system implementing the first embodiment of the present invention can be implemented by a training module 510, a service module 530, and a matching module 540.

[0146] For each feature, multiple images 512, grouped into categories including similar and dissimilar images based on their vector definition 550, are fed into the training module 510.

[0147] In the training module, the one or more image classification classifiers are implemented as Siamese neural networks. In an exemplary embodiment, the colored hat pattern, the colored body pattern, and the colored sleeve pattern are parameter-defined features or feature groups used in such an arrangement in the exemplary embodiment to identify jockeys in horse racing.

[0148] Similarly, after training, one or more trained image classifiers 520 are obtained and saved.

[0149] For the service period, for each feature or feature group, the service module 530 uses a trained image classifier 520 to predict the score of each unique feature or feature group.

[0150] If necessary, the service module uses an object detector to extract images representing features for processing.

[0151] Similarly, the matching module 540 is implemented to calculate and identify the set of parameters corresponding to the features or feature groups that have the highest or best combination score according to the matching algorithm.

[0152] Similarly, based on the identified feature categories, the matching module 540 predicts the identity of the most likely participant in image 532 among the participants in the source images.

[0153] Figure 6 illustrates a flowchart 600 of the definition portion of the training phase of the second embodiment of the present invention, and how this phase prepares for the training of the A model and the B model used in various arrangements of the second embodiment discussed further herein.

[0154] In step 602, a collection of participant badges for various completed events of the same type (such as competitions or contests) is gathered.

[0155] Similar to step 202, the corpus of images of known participant emblems should cover images of all distinguishing features (such as patterns, shapes, and colors) in the common / public parts of the emblems of participants of a certain type of event.

[0156] Advantageously, and similarly to the advantages of the first embodiment, it is not necessary to have images of all participants' emblems in the corpus in the specific and unique combination that these participants will be wearing for the upcoming event.

[0157] For the training phase, and within both the definition and training modules, it is necessary to ensure that the features and feature parameter values ​​in the images depicting participant badges from completed events include images with all the individual features and parameter values ​​of participant badges that will appear in upcoming events. For example, if the features of a participant badge in an upcoming event include a herringbone sleeve pattern and yellow and red colors, then the participant image set does not need to include images of a specific combination of sleeve pattern and color (i.e., a herringbone sleeve with yellow and red colors), and can simply include images of the herringbone sleeve pattern and yellow and red sleeves.

[0158] The total number of images required and the number of images for each feature or feature group will depend on the training method, the design of the common / public parts of the participant badges, and the type of design of the common / public parts of the participant badges.

[0159] In the case of classifying the identity of jockeys or cyclists, approximately 1000 classes of jockey or cyclist badges and approximately 1000 images of each unique badge, including unique combinations of feature parameter values, are collected. However, the accuracy of the model will increase with a higher number of images of each class of badges used for training. As in the first embodiment, it is required that the badge of each participant in the event should be distinguishable by unique features or feature groups.

[0160] In step 604, depending on the type of emblem design, one or more features or groups of features may be determined for identifying the emblem. In the example of identifying a jockey, features that may be used to identify a participant include hat pattern, hat color, body pattern, body color, sleeve pattern, and sleeve color. Depending on the horse's attire, the horse mask and saddle may also include identifying features of the emblem defined for that event type.

[0161] Alternatively, if participants are in a cycling race, they can use a different set of unique features, such as helmet pattern, helmet shape, helmet color, body pattern, body color, shorts pattern, and shorts color.

[0162] In step 606, similar to the first embodiment of this disclosure, a logo scheme is defined based on the parameter values ​​of the features.

[0163] Figures 12A to 12C and Figure 13 Examples of parameter values ​​for the characteristics of the jockey's badge—hat pattern, body pattern, sleeve pattern, and color—are shown. These are similar to the characteristics mentioned in step 204. Figure 13 The example depicts parameter values ​​for color features.

[0164] In step 608, the image corpus and the corresponding parameter values ​​of the image (i.e., the parameter definition of the image) are stored in the definition module.

[0165] In the second embodiment of this disclosure, and in the second part of the training phase, "Model A" and "Model B" are trained in steps 700 and 800 to perform the key functions; the training of Model B depends on the output vector of Model A as described herein.

[0166] Figure 6B A schematic block diagram of the training phase of the second embodiment of the present invention is shown, as well as a system for implementing the training phase of the second embodiment of the present invention, implemented by model A training module 630 and model B training module 650.

[0167] The A-model training module 630 is configured as a Siamese neural network with identically configured classifiers 632a, 632b, and 632c. These classifiers are trained using tuples of vectors 654, 656, and 658 representing the encoding parameter definitions, as described in step 700 below. After satisfactory training, the trained A-model classifier 636 is obtained and saved.

[0168] The B-model training module 650 is advantageously configured as another twin neural network of image classification classifiers 652a, 652b, 652c with the same configuration, and is trained by anchor vector 654 generated from the vector of encoded parameter values ​​by the trained A-model classifier 636 and tuples of corresponding images 664, 666, as further described herein.

[0169] After satisfactory training, the trained B-model image classifier 660 is obtained and saved.

[0170] like Figure 7 As depicted in flowchart 700, during the training phase of the training module, model A is trained according to the parameter definition such that, once trained, it has a function configured to separate dissimilar parameter values ​​of features far apart in the solution space. Next, as depicted in the steps of flowchart 800, the output of model A is used to train model B to map dissimilar images to classifiers that are spatially far apart.

[0171] The training process for "Model A" and "Model B" is further described below.

[0172] Figure 7 A flowchart 700 depicts the first part (definition module) of the training phase of an arrangement according to a second embodiment of the present invention.

[0173] In step 702, multiple preparation steps can be performed sequentially or in parallel without limitation. In step 702a, based on the relevant emblem scheme stored in the definition module in step 608, parameter definitions for the emblems representing each participant in the upcoming event are selected and loaded.

[0174] In step 702b, the parameter values ​​for each feature or feature group of the parameter definition corresponding to the participant badge are encoded. The parameter values ​​can be encoded using one-hot encoding. Alternatively, parameter values ​​representing color can be encoded as RGB values ​​instead of using one-hot encoding. For example, the parameter values ​​{no pattern, yellow, none, herringbone, yellow, red, V-shape, silver, red} of the participant badge parameter definition for a specific jockey in a particular race (e.g., Mr. Li in race 6 at Happy Valley) are encoded as a vector.

[0175] In the arrangement of the second embodiment of this disclosure, the parameter values ​​of all features of the participant's badge can be directly encoded for training, that is, by grouping all features into one group instead of separating different features / feature groups, as described above for the example jockey Mr. Li {no pattern, yellow, none, herringbone, yellow, red, V-shape, silver, red}.

[0176] Alternatively, in another arrangement of the second embodiment, the parameter definition can be decomposed into features or feature groups to improve the efficiency of one-hot encoding and training.

[0177] For example, if features are grouped by sleeve pattern and sleeve color, hat pattern and hat color, and body pattern and body color, then the above parameter values ​​defined for the example jockey Mr. Li can instead be defined as three separate feature groups: {no pattern, yellow, none}, {herringbone, yellow, red}, and {V-shape, silver, red}.

[0178] In step 702c, the dataset can be split into a training dataset and a testing dataset. A common approach is to randomly split the entire dataset into a 9:1 or 8:2 ratio for training and testing.

[0179] In step 702d, a classification classifier is defined. The classification classifier can be a custom convolutional neural network, or any of the existing neural networks used for image classification (such as Inception V4, VGG16, and ResNet).

[0180] In step 704, for each feature or feature group, a Siamese neural network is defined that utilizes a fully connected dense neural network for classifying the classifier.

[0181] As is known in the art, a Siamese neural network is a network with two or more identical classifiers that share weights and bias parameters. The number of classifiers to be configured in a Siamese neural network depends on the image category representing unique combinations of features and feature groups, the training method, and the loss function. The weights and bias parameters are shared among the two or more classifiers and are adjusted according to the same loss function following the connection function.

[0182] For each feature or feature group, in an exemplary arrangement according to the second embodiment of this disclosure, three inputs are fed into the Siamese neural network. These three vector inputs are an anchor vector, a similar vector, and a dissimilar vector, such as... Figure 6B What is depicted.

[0183] Anchor vector 654 is the encoded parameter value representing the parameter definition of the participant's emblem. Therefore, in the example of Mr. Li provided earlier, the value discussed was encoded in vector format using appropriate encoding (such as one-hot encoding and RGB value encoding) without restriction.

[0184] The similarity vector 656 is the anchor vector plus noise of RGB values ​​randomly injected into the color parameter value, or if the color feature is encoded with RGB values ​​or similar values, it is a vector belonging to the same category as the anchor vector, or a combination of the above methods used to form the similarity vector.

[0185] Dissimilar vector 658 should be a vector of a different class from the anchor vector, or a vector encoded by dissimilar parameter values ​​and having parameter values ​​different from the anchor vector.

[0186] In step 706, for each feature or feature group, the batch size, number of epochs, appropriate loss function, and other hyperparameters of the classification classifier defined in step 704 are configured with the same parameters.

[0187] A suitable loss function is one that adjusts similar vectors (e.g., anchor vectors and similar vectors) to be closer to each other in the output space of the Siamese network's image classifier and / or adjusts dissimilar vectors (e.g., anchor vectors and dissimilar vectors) to be further apart, such as a triple loss function, a circular loss function, or a contrastive loss function.

[0188] In step 708, the training process of the Siamese neural network begins, and the parameters of the Siamese neural network are updated through backpropagation. The model will converge after several epochs of training. Advantageously, if the classification classifier of model A is a custom neural network rather than an image classification classifier, then the Siamese neural network requires less training time to converge, and the Siamese neural network converges after 10 to 20 epochs.

[0189] In step 710, Euclidean distance is calculated to evaluate the distance between vectors of all the participants' badges. The model is considered converged if the Euclidean distance between vectors encoding dissimilar parameter values ​​is significantly greater than the Euclidean distance between vectors encoding similar parameter values. In the exemplary configuration using the triple loss function, the α value is an indicator of the Euclidean separation distance.

[0190] In step 712, for each feature or feature group, the A model classifier is obtained and saved, that is, the corresponding trained classification classifier of the Siamese neural network.

[0191] Figure 8 A flowchart 800 illustrates the second part of the training period according to a second embodiment of the present invention.

[0192] In step 802, the characteristic and characteristic group parameter values ​​of each known participant's emblem for the upcoming event are collected.

[0193] In step 804, the parameter definition (parameter value group) of each of the emblems of the known participants in the upcoming event, which was stored in the definition module in step 608, is loaded.

[0194] In step 806, for each feature or feature group, anchor vectors used as inputs to the B model are computed by passing the parameter definitions of each badge of a known participant in the upcoming event to the trained A model classifier obtained in step 712. The parameter values ​​of these input participant badges are encoded in the same manner as in the training of the corresponding A model described herein.

[0195] Therefore, the anchor vector of each participant's emblem in the upcoming event is the output of a trained A-model for that event type, which is defined by the parameters of each emblem of each known participant in the upcoming event.

[0196] Advantageously, the calculation of all anchor vectors of the participant's emblem can be performed in real time or near real time.

[0197] In step 808, the dataset is split into a training dataset and a test dataset. A common approach is to randomly split the entire dataset into a 9:1 or 8:2 ratio for training and testing.

[0198] In step 810, an image classification classifier is defined. The image classification classifier can be a custom convolutional neural network or one of existing neural networks used for image classification (such as Inception V4, VGG16, and ResNet). If an image classification classifier is used in model A, the image classification classifier used in model B can be the same type of image classification classifier defined for model A.

[0199] In step 812, for the same features or feature groups corresponding to the A model trained for each feature or feature group, a Siamese neural network for image classification classifiers is defined. A Siamese neural network is a network of two or more identical image classification classifiers with shared weights and bias parameters. The number of image classification classifiers to be configured in the Siamese neural network depends on the training method and loss function.

[0200] In step 814, the anchor vectors calculated from model A and the corresponding known similar and / or dissimilar participant images are loaded.

[0201] For training the B model, load the corresponding similar image 664 with the same parameter values ​​as the anchor vector 654.

[0202] Preferably, if a triple loss function is used, the Siamese neural network can accept dissimilar images 666 with parameter values ​​that are not similar to or different from those of the anchor vector 654.

[0203] Advantageously, additionally inputting random images dissimilar to anchor vector 654 and similar images 664 into the Siamese neural network will improve the effectiveness of the image classifier.

[0204] In step 816, for each feature or feature group, the batch size, number of epochs, appropriate loss function, and other hyperparameters of the image classification classifier are defined in step 810.

[0205] A suitable loss function is one that adjusts similar vectors (e.g., anchor vectors and positive vectors) to be close to each other in the output space of the Siamese network's classifier and / or adjusts dissimilar vectors (e.g., anchor vectors and negative vectors) to be far apart, such as a triple loss function, a circular loss function, or a contrastive loss function. Model A is trained to map similar anchor vectors to be close to each other or clustered in the output space of the image classifier, and to map dissimilar anchor vectors to be far apart.

[0206] Model B is trained using anchor vectors generated from Model A, and maps similar images to be close to each other or clustered in the output space of the image classifier, while mapping dissimilar images to be far apart, making the image classifier perform well in distinguishing between similar and dissimilar images.

[0207] In step 818, the training process of the Siamese neural network begins, updating its parameters through backpropagation, and the model can converge after several hundred epochs of training. In the example, the Siamese neural network for the image classifier converges after 100 to 300 epochs.

[0208] In step 820, the B-model classifier for the feature or feature group is obtained and saved. Steps 806 to 820 are repeated for each feature or feature group.

[0209] Advantageously, a second embodiment of the invention can be trained using a feature set that includes all features corresponding to the participant's emblem.

[0210] Model A and Model B are pre-trained models, and advantageously, they do not require special training for specific events.

[0211] Advantageously, once the A-model classifier and the B-model classifier are trained, they can be further configured for all the same types of events, including the applicable feature parameter value categories.

[0212] Figure 9 A flowchart 900 illustrates the configuration process of Participant A Model according to a second embodiment of the present invention.

[0213] In step 902, the trained A model obtained in step 712 is used.

[0214] In step 904, the emblem parameter definition for each participant's emblem in the event is obtained.

[0215] The participant badges that will appear in the upcoming event can be either seen participant badges or unseen participant badges, provided that unseen participant badges are covered by the features of the badge scheme and can be encoded by the badge scheme.

[0216] In step 906, each parameter definition of the participant's emblem is encoded using the encoding method employed in training Model A. This encoding method can use a combination of one-hot encoding or RGB value encoding without limitation.

[0217] In step 908, by inputting the encoded parameter definition of each participant's emblem, the vector Kp of each parameter definition of the known participant's emblem is calculated using the A model from step 712.

[0218] In step 910, the vector Kp defined for each parameter of the participant's emblem is saved and referred to as the anchor vector for that feature / feature group of that participant. Advantageously, the steps of the configuration process using Model A can be performed in real time using a pre-trained model.

[0219] Figure 10 The flowchart 1000, which describes the steps of a process for running a B-model to identify the participant badges of an event according to a second embodiment of the present invention, is described.

[0220] In step 1002, for each feature or feature group, the trained B model obtained from step 820 is used.

[0221] In step 1004, frame images or pictures extracted from video, image streams, or broadcasts are analyzed.

[0222] In step 1006, object detector model 1 is applied to locate the coordinates (bounding boxes) of each participant in the frame from step 1004.

[0223] If models A and B are trained on features or feature groups, then appropriate object detectors (such as object detector models p1 to pN, i.e., the same object detectors for image preprocessing disclosed herein with reference to the first embodiment) are used to crop regions of the participant images that represent emblem portions having features or feature groups.

[0224] In step 1008, images of participants located in the frame are captured by object detector model 1.

[0225] In step 1010, a vector Xp representing each image or image region of the participant in step 1008 is obtained by inputting the participant's image or image region into the B model.

[0226] In step 1012, the Euclidean distance of the vector of the image of the unknown (i.e., unidentified) participant in the event is measured relative to the anchor vector Kp, which is generated based on the parameter definition of each known participant in the event and calculated from model A.

[0227] In step 1014, based on the Euclidean distance obtained from step 1012, the parameter definition and participant identity of the closest vector (i.e., the vector with the shortest distance between vector Xp and the known anchor vector Kp) are assigned.

[0228] If Model A and Model B are trained with multiple features or feature groups, a matching algorithm (such as the minimum or weighted sum of all Euclidean distances of all features or feature groups) is applied to determine the most likely clothing parameter definitions for the participants.

[0229] In step 1016, the status of the participant identified in step 1014 is assigned as "known".

[0230] In step 1018, based on the database of participant badges and corresponding participant identities, the identities of the participants present in the image from step 1006 are returned.

[0231] Figure 11 A schematic block diagram of the service phase of the arrangement according to a second embodiment of the present invention is shown.

[0232] As depicted, the system implementing the second embodiment of the present invention can be implemented by a trained A-model image classifier 1160, a trained B-model image classifier 1162, and a matching module 1140.

[0233] The definition module stores the parameter values ​​of each known participant's emblem, which are required to generate the anchor vector 1122 of each known participant in the event based on the encoded parameter definition of each participant's emblem 1110.

[0234] During the service, images 1135 of unidentified participants are fed into a trained B-model image classifier 1162 to obtain the vector of the unknown participant.

[0235] Matching module 1140 calculates the Euclidean distance of the unknown participant vector 1138 relative to each of the anchor vectors 1122 of the known participants in the event, in order to determine the participant identity of the unknown participant based on the Euclidean distance 1130.

[0236] The matching module can optionally mark the identified participants as "known" and associate the participants' identities, and then the participants' identities can be displayed on an image or provided to the user through channels, etc.

[0237] Furthermore, advantageously, a combination of the first and second embodiments of this disclosure can be used to improve the overall accuracy of participant identification in an event.

[0238] In the example, the methods disclosed herein in the first and second embodiments are combined, and if the minimum Euclidean distance d of the image of the participant's emblem or its region is assessed by the second embodiment to be below a threshold (determined experimentally), that is, when the confidence of the prediction is high, the prediction of the second embodiment can be used to improve the accuracy of the prediction of the first embodiment.

[0239] In alternative examples of the combined use of the methods disclosed in the systems and methods of this disclosure, different combinations of features and feature groups with the same set of parameter values ​​may also be used to improve overall accuracy. A first classifier is trained to identify participants based on a single feature using a portion of an image, while a second classifier is trained to identify participants based on feature groups using images of participants that include the features. The parameter definitions and detectable parameter values ​​of the features are compatible between the first and second classifiers. The first and second classifiers can be used in combination to provide higher confidence in predictions of participants in (multiple) unknown images.

[0240] Figures 12A to 12C and Figure 13The pattern and color features of the insignia of the event participants are shown respectively. In an embodiment of the invention, the pattern features are a hat pattern, a body pattern, and a sleeve pattern, which are respectively located in... Figures 12A to 12C Depicted in the middle. As depicted, from left to right, Figure 12A The top row depicts exemplary jockey hat patterns in herringbone, ring, and ring-2 color schemes; the middle row shows ring tiles, plain patterns, and stars; the bottom row shows stripes, multiple stripes, and striped tiles. Similarly, in Figure 12B The sweatshirt pattern depicted has the following elements: the top row contains checks, multiple diamonds, and disc patterns; the second row contains two colors, multiple rings, and large dots; the third row is plain, with dots and diagonal stripes; and the bottom row contains three colors, three diamonds, and a V-shaped pattern. Figure 12C The depicted sleeve pattern shows herringbone, multiple herringbone, and diamond patterns in the top row; the middle row is colored, tiled in rings, and plain; and the bottom row is dots and stripes. It should be understood that these are merely exemplary and other patterns may be used without departing from this disclosure.

[0241] Figure 13 Various exemplary color blocks are depicted, and these color blocks can be applied independently. Figures 12A to 12C The design may include patterns depicting the hat, body, and sleeves, or, appropriately, other parts of the emblem. For any given pattern feature, each individual pattern should be unique within the set of patterns. Preferably, each individual pattern is distinguishable from the others; and similar patterns may be used, such as single stripes, tricolor stripes, and striped tilings or rings, two-color rings, and ring tilings.

[0242] Figure 14 Several exemplary parameter definitions for the jockey's silk attire are shown, along with a visual representation of the overall emblem and its individual parts. Reference Figure 14 In the first example, parameter definitions can be represented as a list of field value pairs.

[0243] The fields and values ​​defined for the parameters remain consistent throughout the definition, model training, and service phases. In this example, Figure 14The parameters for the three jockey silk outfits depicted are defined as follows: {ID: A024, sleeve pattern: plain, sleeve base color: yellow, sleeve secondary color: none, hat pattern: herringbone, hat base color: yellow, hat secondary color: red, body pattern: V-shape, body base color: silver, body secondary color: red}; {ID: A044, sleeve pattern: plain, sleeve base color: orange, sleeve secondary color: none, hat pattern: stars, hat base color: gold, hat secondary color: orange, body pattern: striped, body base color: royal blue, body secondary color: gold}; and {ID: A049, sleeve pattern: herringbone, sleeve base color: red, sleeve secondary color: pink, hat pattern: plain, hat base color: pink, hat secondary color: none, body pattern: multiple diamonds, body base color: pink, body secondary color: red}.

[0244] result

[0245] The technical solution described in the first embodiment is implemented using emblem definition and trained using an image dataset that includes a collection of jockey silk outfits and jockey attire from the race.

[0246] Approximately 2000 or more images were collected, each featuring a jockey's badge associated with a horse race. The features defined in the badge scheme within the definition module are hat pattern, hat color, body color, sleeve pattern, and sleeve color, and the images of these features were cropped using object detector 1. In one example executed, a custom convolutional neural network was used as the classifier.

[0247] The dataset was randomly split into training and testing datasets in a 9:1 ratio. For each feature, the training parameters of the image classification classifier were defined using a contrastive loss function. The image classification classifier was trained and obtained after 20 epochs. The system running the training module had 64 CPU cores and 4 GPU accelerator cards, each with 32GB of GPU RAM.

[0248] The service method according to a first embodiment of the present invention applies an image classification classifier to 450 competition events and compares the predictions of the service method with the basic facts.

[0249] Table 1 shows the prediction results of participant identities using the first arrangement of the first embodiment. To measure the results, precision and recall were calculated based on the predictions and basic facts.

[0250] Accuracy=(True Positives) / (True Positives+False Positives)

[0251] Recall rate = (True Positives) / (True Positives + False Negatives)

[0252] The average precision obtained was 97.92%, with a standard deviation of 2.67%. The average recall was 87.87%, with a standard deviation of 4.3%.

[0253] These parameters indicate that the trained model maintains high accuracy at an acceptable level of recall.

[0254] count percentage >90% accuracy 441 98％ >92% accuracy 431 95.78％ >95% accuracy 412 91.56％ >98% accuracy 302 67.11％ >80% recall rate 427 94.89％ >70% recall rate 450 100％

[0255] Table 1

[0256] The technical solution described in the second embodiment is implemented using an exemplary emblem definition and trained on an image dataset including a set of jockey outfits and jockey emblems from races. The feature groups defined in the jockey emblem scheme for horse racing by the user configuration of the definition module are hat patterns with hat colors, body patterns with body colors, and sleeve patterns with sleeve colors. Approximately 1000 images of 1000 unique jockey outfits and each type of jockey outfit were collected. The images of the features were cropped by an object detector 1. An Inception V4 neural network was used as the image classification classifier. For each feature group, the training parameters of the image classification classifier were defined using a triple loss function. The image classification classifier was trained and obtained after 300 epochs. The system running the training module has 64 CPU cores and 4 GPU accelerator cards, each with 32GB of GPU RAM.

[0257] Images of 935 unique jockey silk outfits were collected, with approximately 1,200 images of each outfit from unseen races, to create a validation dataset for testing.

[0258] Group 1 of the validation dataset (seen patterns) includes 569 combinations of seen emblem patterns from the training dataset.

[0259] Group 2 of the validation dataset (unseen patterns) includes 366 combinations of seen emblem patterns from the training dataset.

[0260] Group 3 of the validation dataset (a mixture of seen and unseen patterns) includes 935 combinations of emblem patterns.

[0261] As described, an image classification classifier is used during the service phase.

[0262] Calculate the N-way accuracy for each group. Random images are selected from the validation datasets of Groups 1, 2, and 3. For each N-way test, 15 images are randomly selected from the dataset. The vector representing the image is computed by Model B and compared with the anchor vector computed by Model A. The recognition of unknown images is based on the minimum Euclidean distance between the vector of the unknown image and the anchor vector according to the above method. 200,000 recognitions have been tested, and the N-way accuracy is calculated based on the count of correct recognitions divided by the total count (i.e., correct_identification_count / total_count).

[0263] The 15-way accuracy for groups 1, 2, and 3 is:

[0264] Group 1 15-way accuracy: 99.75%

[0265] Group 2 15-way accuracy: 84.75%

[0266] Group 3 15-way accuracy: 96.85%

[0267] refer to Figure 16 A schematic diagram of a system 1600 implementing the first and second embodiments of the present invention is shown below. The system implementing the training module, service module, and matching module includes one or more computers having one or more CPUs 1604, memory 1601, input interface 1602 and output interface 1603, controller 1605, monitor 1606, and network interface 1607. As those skilled in the art will understand, GPUs, application-specific integrated circuits (ASICs), and other hardware may be added to system 1600 to improve training speed.

[0268] It should be understood that the above embodiments are described by way of example only. Many variations are possible without departing from the scope of the invention as defined in the appended claims. For clarity, in some instances, the technology may be presented as comprising various functional blocks, which include steps or routines in a method implemented in software or a combination of hardware and software.

[0269] The methods described in the examples above can be implemented using computer-executable instructions stored in or otherwise made available from a computer-readable medium. Such instructions may include, for example, instructions and data that cause or otherwise configure a general-purpose computer, special-purpose computer, or special-purpose processing device or unit to perform a particular function or group of functions.

[0270] Some computer resources used may be accessible via a network. Computer-executable instructions may be, for example, binary files, intermediate format instructions such as assembly language, firmware, or source code. Examples of computer-readable media that may be used to store instructions, information used, and / or information created during the process of the method according to the described examples include disks or optical discs, flash memory, Universal Serial Bus (USB) devices with non-volatile memory, networked storage devices, etc.

[0271] Devices implementing the methods according to these disclosures may include hardware, firmware, and / or software, and may take any of a variety of form factors. Typical examples of these form factors include laptop computers, smartphones, minicomputers, personal digital assistants, tablet computers, etc. The functionality described herein may also be implemented in peripheral devices or add-on cards. As a further example, such functionality may also be implemented on a circuit board of different chips, or on different processes that may be executed in a single device.

[0272] Instructions, media for transmitting such instructions, computing resources for executing such instructions, and other structures for supporting such computing resources are means for providing the functionality described in these disclosures.

[0273] Although various examples and other information are used to interpret aspects within the scope of the appended claims, no limitation on the claims should be implied based on specific features or arrangements in such examples, as those skilled in the art will be able to derive a wide variety of implementations from these examples. Furthermore, and although a subject matter may have been described in language specific to structural features and / or method steps, it should be understood that the subject matter defined in the appended claims is not necessarily limited to these described features or actions. For example, such functionality may be distributed differently or performed in components other than those identified herein. Rather, the described features and steps are disclosed as examples of components of systems and methods within the scope of the appended claims.

Claims

1. A computer-based method for identifying at least one participant among a plurality of participants in one or more images of an image stream of an event, wherein, Each of the participants has a unique instance of a public emblem associated with it, and the method includes: Receive specified parameter values ​​for one or more features or groups of features that constitute the emblem of each participant in the current event; Store the relationship between a unique combination of specified parameter values ​​representing the features or feature groups of the emblem and the identity of each participant in the current event; After the current event begins, extract one or more image regions from one or more images in the image stream of the current event, each region depicting at least a portion of the emblem of one participant in the event; and The identities of participants in the event are predicted from one or more images in the image stream of the event using the following steps: (i) Determine a matching score between the detection parameter value of each feature or feature group of the corresponding portion of the emblem of each participant in the event and a stored specified parameter value; wherein the determination is performed by using a trained image classifier for each of the one or more image regions in the one or more images of the event, and for each feature or feature group of the emblem of each participant depicted therein; and (ii) The identity of a participant is determined by the best corresponding overall score that matches the detection parameter values ​​of the features or feature groups depicted in the one or more images with the stored specified parameter values ​​of the features or feature groups of the emblem of each participant in the event.

2. The computer-based method for identifying at least one participant among multiple participants in one or more images in an image stream of an event, as described in claim 1, wherein, The features or groups of features of the public emblem include one or more of the following: pattern, shape, and color.

3. The computer-based method for identifying at least one participant among multiple participants in one or more images in an image stream of an event, as described in claim 1, wherein, The trained image classifier is selected from the group of neural network classifiers that include: Inception V4 neural network, VGG16 neural network, ResNet neural network, or convolutional neural network.

4. The computer-based method for identifying at least one participant among multiple participants in one or more images in an image stream of an event, as described in claim 1, wherein, The scores of the detection parameter values ​​are output directly from the trained image classifier.

5. The computer-based method for identifying at least one participant among multiple participants in one or more images of an image stream of an event, as described in claim 1, wherein, The score of the detection parameter value is calculated from the output of the trained image classifier by the classification classifier.

6. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, claim 4, or claim 5, wherein, The optimal corresponding overall score for the parameter values ​​is determined by multiplying the scores determined for the parameter values ​​of all features of each part of the public emblem associated with each participant in the event.

7. The computer-based method for identifying at least one participant among multiple participants in one or more images in an image stream of an event, as described in claim 1, wherein, The specified parameter value is a vector representation of the parameter values ​​of a feature or feature group of a part of the emblem associated with the participant, and the vector representation is generated through the following steps: (i) Select one or more second trained classifiers for the feature or feature group; wherein each second trained classifier is trained to map similar vectors to be spatially relatively close in the output space of the second trained classifier, and to map dissimilar vectors to be spatially relatively far apart in the output space of the second trained classifier; and (ii) Provide the second trained classifier selected for the feature or feature group with corresponding encoded parameter values ​​for one or more portions of the common emblem of the participants of the event to generate the vector representation.

8. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 7, wherein, The second trained classifier is trained using a classifier configured as a Siamese neural network. The inputs to the Siamese neural network of the classifier include: Anchor vectors for encoding parameter values; Similar vectors whose encoded parameter values ​​are similar to the anchor vector, including the anchor vector with noise added to the encoded parameter values, or vectors obtained from combinations of parameter values ​​that are the same as the anchor vector; Dissimilar vectors whose parameter values ​​are not similar to the anchor vector, including vectors that have different combinations of parameter values ​​compared to the anchor vector.

9. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 8, wherein, One or more of the parameter values ​​of the encoding parameter value are encoded using one-hot encoding.

10. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 8, wherein, One or more of the parameter values ​​of the encoding parameter value are encoded using RGB values.

11. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 8, wherein, The loss function for the Siamese neural network used in the classifier is selected from either a contrastive loss function or a triple loss function.

12. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, The trained image classifier is trained using an image classifier configured as a Siamese neural network; Wherein, the loss function of the Siamese neural network used for the trained image classifier is a contrastive loss function; and The input to the Siamese neural network is either an anchor image and a tuple of images with parameter values ​​similar to those of the anchor image, or an anchor image and a tuple of images with parameter values ​​dissimilar to those of the anchor image.

13. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, The trained image classifier is trained using an image classifier configured as a Siamese neural network; Wherein, the loss function of the Siamese neural network used for the trained image classifier is a triple loss function; and The input to the Siamese neural network is a tuple consisting of an anchor image, an image with parameter values ​​similar to those of the anchor image, and an image with parameter values ​​dissimilar to those of the anchor image.

14. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, The trained image classifier is trained using an image classifier configured as a Siamese neural network; Wherein, the loss function of the Siamese neural network used for the trained image classifier is a contrastive loss function; and The input to the Siamese neural network is either a vector representation of a specified parameter value and a tuple of an image with parameter values ​​similar to the specified parameter values, or a vector representation of a specified parameter value and a tuple of an image with parameter values ​​dissimilar to the specified parameter values.

15. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, The trained image classifier is trained using an image classifier configured as a Siamese neural network; Wherein, the loss function of the Siamese neural network used for the trained image classifier is a triple loss function; and The input to the Siamese neural network is a vector representation of specified parameter values, an image with parameter values ​​similar to the specified parameter values, and a tuple of images with parameter values ​​dissimilar to the specified parameter values.

16. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, For each participant’s features or feature groups, the score comparison between the detection parameter value of each participant in the current event and the stored specified parameter value is determined using the Hungarian algorithm.

17. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, For at least some of the features or feature groups of each participant, the score for comparing the detection parameter value of each participant in the current event with the stored specified parameter value is determined based on the distance between the vector representation of the detection parameter value of the feature or feature group of each participant in the current event and the corresponding vector representation of the stored specified parameter value.

18. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, The public emblem associated with the participant includes items of clothing worn by the participant and / or parts of accompanying entities.

19. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 18, wherein, The characteristics of the participant's emblem and / or accompanying entity include patterns, shapes, and colors.

20. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 18, wherein, The public emblem includes any one or more of the hat, body, sleeves, shorts, and boots.

21. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 18, wherein, The accompanying entities of the participants include any one or more entities selected from the group consisting of: horses, vehicles, bicycles, and motorcycles.

22. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 1, wherein, The trained image classifier is selected from the group consisting of: Inception V4 neural network, VGG16 neural network, ResNet neural network, or convolutional neural network.

23. The computer-based method for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, as described in claim 7, wherein, The second trained classifier is selected from the group consisting of: a custom neural network or any one of the Inception V4 neural network, VGG16 neural network, ResNet neural network, and convolutional neural network.

24. A computer-based method for training a system for identifying at least one of a plurality of participants in one or more images in an image stream of an event, wherein, Each of the participants has a unique instance of a public emblem associated with it, and the method includes: Receive a set of images or portions thereof depicting emblems of participants in an ended event; wherein the depicted emblems have detectable parameter values ​​of features or feature groups of common emblems of participants in the ended event; wherein the set of images or portions thereof is labeled according to the detectable parameter values ​​of the emblems of the participants depicted therein; and A subset of training images is generated from the set of the images or portions thereof, having specified parameter values ​​for one or more features or feature groups that are associated with a badge for each participant in the current event; At least one classifier is trained on the subset of training images for each feature or each group of features to score an image or a portion thereof for a specified parameter value of that feature or group of features. The training includes providing a corresponding classifier with input including a corresponding subset of training images for each feature or each group of features; and iteratively updating the classifier for each feature or each group of features until each classifier reaches a predetermined scoring accuracy.

25. The computer-based method according to claim 24, wherein, The collection of images of participants in the concluded event, or portions thereof, does not include images with all parameter values ​​that possess the characteristics of the participants' emblems in the event.

26. The computer-based method according to claim 24, wherein, The collection of images or portions thereof of participants in the concluded event includes multiple images of features of the emblem portion, the multiple images containing all parameter values ​​of the features of the emblem of the participants in the event.

27. The computer-based method according to claim 24, wherein, The at least one classifier is an image classifier selected from the group of neural network classifiers including: Inception V4 neural network, VGG16 neural network, ResNet neural network, or convolutional neural network.

28. The computer-based method according to claim 24, wherein, The training is the training of two or more classifiers configured as a Siamese neural network, which uses a contrastive loss function, a triple loss function, or a cross-entropy loss function.

29. The computer-based method according to claim 28, wherein, The input to the Siamese neural network of the classifier includes either a tuple of anchor images and similar images, or a tuple of anchor images and dissimilar images.

30. The computer-based method according to claim 28, wherein, The input to the Siamese neural network of the classifier includes tuples of anchor images, similar images, and dissimilar images.

31. The computer-based method according to claim 28, wherein, The input to the Siamese neural network of the classifier includes anchor vectors, tuples of similar and dissimilar images.

32. The computer-based method according to claim 24, wherein, The collection of images is generated after performing one or more image preprocessing steps, which are selected from the group including resizing and augmenting the acquired images.

33. The computer-based method according to claim 31, wherein, The anchor vector is generated through the following steps: (i) Select one or more second trained classifiers for the feature or feature group; wherein each second trained classifier is trained to map similar vectors to be spatially relatively close in the output space of the second trained classifier, and to map dissimilar vectors to be spatially relatively far apart in the output space of the second trained image classifier; and (ii) Provide a second trained classifier selected for the feature or feature group with corresponding encoded parameter values ​​for features or feature groups corresponding to one or more portions of the public emblem of the participants in the event to generate the anchor vector.

34. The computer-based method according to claim 33, wherein, The second trained classifier is trained using a classifier configured as a Siamese neural network. The inputs to the Siamese neural network of the classifier include: A parameter vector that encodes parameter values; A similar vector whose encoded parameter values ​​are similar to the parameter vector, the similar vector including the parameter vector after adding noise to the encoded parameter values, or a vector obtained from a combination of parameter values ​​that are the same as the parameter vector; Dissimilar vectors whose parameter values ​​are not similar to the parameter vector, including vectors that have different combinations of parameter values ​​compared to the parameter vector.

35. The computer-based method according to claim 34, wherein, One or more of the parameter values ​​in the parameter vector are encoded using one-hot encoding.

36. The computer-based method according to claim 34, wherein, One or more of the parameter values ​​in the parameter vector are encoded using RGB values.

37. A computer-readable medium comprising instructions that, when executed by a processor, cause the processor to perform the method as described in any of the preceding claims.

38. A system for identifying at least one participant among a plurality of participants in one or more images in an image stream of an event, wherein, Each of the participants has a unique instance of a public emblem associated with it, and the system includes an image processing module configured to perform the method as described in any one of claims 1 to 23.

39. The system of claim 38, further comprising a training module configured to perform the method of any one of claims 24 to 36.

Images