An object recognition method, apparatus, computer device, and storage medium

Abstract

Embodiments of the present application disclose an object recognition method and device, computer equipment and a storage medium. The embodiments of the present application can be applied to the field of automatic driving or transportation. The embodiments of the present application can obtain object sample data in a target object recognition scenario and enhanced object sample data corresponding to the object sample data. The enhanced object sample data is subjected to feature extraction by a plurality of different object recognition models to obtain homologous object features. Based on the homologous object features, a robustness measurement score of each object recognition model is obtained. Based on reference object feature information of each object recognition model, an accuracy measurement score corresponding to each object recognition model is obtained. According to the accuracy measurement score and the robustness measurement score of each object recognition model, a target object recognition model suitable for the target object recognition scenario can be accurately selected from the plurality of object recognition models to perform object recognition through the target object recognition model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to an object recognition method and device, computer equipment and storage medium. BACKGROUND

[0002] Object recognition is a kind of biometric technology based on object feature information for identity recognition. It is a series of related technologies for automatically detecting and determining objects in images or video streams collected by cameras or cameras, and then performing object recognition on the detected objects. With the continuous development of computer technology and artificial intelligence technology, object recognition technology has been widely applied in many application scenarios. For example, an object recognition model can be trained by artificial intelligence technology, and then object detection can be performed by using the object recognition model. However, since the distribution of object image data types in different application scenarios is inconsistent, it is required that the deployed object recognition model is adapted to object images in different scenarios. SUMMARY

[0003] Embodiments of the present application provide an object recognition method and device, computer equipment and storage medium, which can accurately select a target object recognition model most suitable for a target object recognition scene from a plurality of object recognition models.

[0004] The embodiments of the present application provide an object recognition method, comprising:

[0005] Obtain object sample data under a target object recognition scene, and perform data enhancement processing on the object sample data to obtain enhanced object sample data;

[0006] Extract features from the enhanced object sample data by using a plurality of different object recognition models to obtain homologous object features extracted by each object recognition model;

[0007] Based on the homologous object features extracted by each object recognition model, obtain a robustness measurement score corresponding to each object recognition model;

[0008] Based on the reference object feature information extracted by each object recognition model for the object sample data, obtain an accuracy measurement score corresponding to each object recognition model;

[0009] According to the accuracy measurement score and the robustness measurement score of each object recognition model, select a target object recognition model suitable for the target object recognition scene from the plurality of different object recognition models, so as to perform object recognition by using the target object recognition model.

[0010] Correspondingly, the embodiments of the present application also provide an object recognition device, comprising:

[0011] a data enhancement unit, configured to obtain object sample data in a target object recognition scenario, and perform data enhancement processing on the object sample data to obtain enhanced object sample data;

[0012] a feature extraction unit, configured to perform feature extraction on the enhanced object sample data by using a plurality of different object recognition models to obtain homologous object features extracted by each object recognition model;

[0013] a first score acquisition unit, configured to acquire a robustness measurement score corresponding to each object recognition model based on the homologous object features extracted by each object recognition model;

[0014] a second score acquisition unit, configured to acquire an accuracy measurement score corresponding to each object recognition model based on reference object feature information extracted by each object recognition model for the object sample data;

[0015] a screening unit, configured to screen a target object recognition model that is suitable for the target object recognition scenario from the plurality of different object recognition models according to the accuracy measurement score and the robustness measurement score of each object recognition model, so as to perform object recognition by using the target object recognition model.

[0016] In an embodiment, the second score acquisition unit can include:

[0017] a first identification subunit, configured to identify the reference object feature information extracted by each object recognition model to obtain an accuracy reference threshold corresponding to each object recognition model;

[0018] a second identification subunit, configured to identify a recall threshold list corresponding to each object recognition model in the reference object feature information of each object recognition model according to label information corresponding to the object sample data;

[0019] a first discrimination subunit, configured to discriminate the recall threshold list corresponding to each object recognition model according to the accuracy reference threshold corresponding to each object recognition model to obtain the accuracy measurement score corresponding to each object recognition model.

[0020] In an embodiment, the first identification subunit can include:

[0021] a combination module, configured to combine the object sample data to obtain an object sample data group;

[0022] a calculation module, configured to calculate object space feature similarity between the object sample data in the object sample data group based on the reference object feature information;

[0023] An identifying module is configured to identify a set of heterogeneous image groups from the set of object sample data according to label information of the object sample data in the set of object sample data.

[0024] A generating module is configured to generate an accuracy reference threshold corresponding to the object recognition model based on object space feature similarity of the set of object sample data in the set of heterogeneous image groups.

[0025] In an embodiment, the determining subunit can include:

[0026] A comparing module is configured to compare the object space feature similarity in the recall threshold list with the accuracy reference threshold to obtain a comparison result.

[0027] A determining module is configured to determine a misjudgment feature similarity from the recall threshold list according to the comparison result.

[0028] A comparison operation module is configured to perform a comparison operation on the number of misjudgment feature similarities and the number of object space feature similarities in the recall threshold list to obtain an accuracy measurement score corresponding to the object recognition model.

[0029] In an embodiment, the first score obtaining unit can include:

[0030] A combining subunit is configured to combine a plurality of homologous object features identified by the object recognition model to obtain a homologous object feature group.

[0031] A calculating subunit is configured to calculate similarity distances between homologous object features in the homologous object feature group.

[0032] A denoising subunit is configured to perform denoising processing on the similarity distances of the homologous object feature group to obtain homologous object feature similarity of the object recognition model.

[0033] A second determining subunit is configured to determine a robustness measurement score corresponding to each object recognition model based on the homologous object feature similarity corresponding to each object recognition model.

[0034] In an embodiment, the data enhancement unit can include:

[0035] An obtaining subunit is configured to obtain a plurality of different data enhancement methods.

[0036] A parameter determining subunit is configured to determine an enhancement parameter corresponding to each data enhancement method.

[0037] A data enhancement subunit is configured to perform data enhancement processing on the object sample data by using a plurality of different data enhancement methods based on the enhancement parameter corresponding to each data enhancement method to obtain enhanced object sample data.

[0038] In an embodiment, the object recognition apparatus can further include:

[0039] an acquisition unit configured to acquire an object recognition model to be trained and object training sample data;

[0040] a feature extraction unit configured to perform spatial feature extraction on the object training sample data by using the object recognition model to be trained, to obtain spatial structure information corresponding to the object training sample data;

[0041] a fully connected mapping unit configured to perform fully connected mapping processing on the spatial structure information corresponding to the object training sample data by using the object recognition model to be trained, to obtain an identification result corresponding to the object training sample data;

[0042] a calculation unit configured to calculate model loss information based on preset label information of the object training sample data and the identification result corresponding to the object training sample data;

[0043] an adjustment unit configured to adjust model parameters of the object recognition model to be trained based on the model loss information, to obtain an object recognition model.

[0044] Embodiments of the present application also provide a computer program product or a computer program, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to perform the method provided in any of the various optional manners of the above aspect.

[0045] Correspondingly, the embodiments of the present application also provide a storage medium, which stores instructions. The instructions are executed by a processor to implement the object recognition method provided in any of the embodiments of the present application.

[0046] The embodiments of the present application can acquire object sample data in a target object recognition scenario and enhanced object sample data corresponding to the object sample data; perform feature extraction on the enhanced object sample data by using a plurality of different object recognition models, to obtain homologous object features; acquire a robustness measurement score of each object recognition model based on the homologous object features; acquire an accuracy measurement score corresponding to each object recognition model based on reference object feature information of each object recognition model; and accurately select a target object recognition model that is suitable for the target object recognition scenario from the plurality of object recognition models according to the accuracy measurement score and the robustness measurement score of each object recognition model, to perform object recognition by using the target object recognition model. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0048] Figure 1 is a scene schematic diagram of an object recognition method provided by an embodiment of the present application;

[0049] Figure 2 is a flow schematic diagram of an object recognition method provided by an embodiment of the present application;

[0050] Figure 3 is still another scene schematic diagram of an object recognition method provided by an embodiment of the present application;

[0051] Figure 4 is still another scene schematic diagram of an object recognition method provided by an embodiment of the present application;

[0052] Figure 5 is still another scene schematic diagram of an object recognition method provided by an embodiment of the present application;

[0053] Figure 6 is still another flow schematic diagram of an object recognition method provided by an embodiment of the present application;

[0054] Figure 7 is a structure schematic diagram of an object recognition apparatus provided by an embodiment of the present application;

[0055] Figure 8 is a structure schematic diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with the drawings in the embodiments of the present application. However, the described embodiments are only some embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] An object recognition method is provided by an embodiment of the present application, which can be executed by an object recognition apparatus. The object recognition apparatus can be integrated in a computer device. The computer device can include at least one of a terminal and a server. That is, the object recognition method provided by an embodiment of the present application can be executed by a terminal, a server, or a terminal and a server capable of mutual communication.

[0058] The terminal can include, but is not limited to, a smartphone, a tablet computer, a notebook computer, a personal computer (PC), a smart home appliance, a wearable electronic device, a VR / AR device, a vehicle-mounted terminal, a smart voice interaction device, and the like.

[0059] The server can be an interworking server or a background server between multiple heterogeneous systems, can also be a standalone physical server, can also be a server cluster or a distributed system composed of multiple physical servers, can also be a cloud server providing basic cloud computing services such as cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, and big data and artificial intelligence platform, and the like.

[0060] It should be noted that the embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, and assisted driving.

[0061] In an embodiment, as Figure 1 The object recognition device can be integrated in a computer device such as a terminal or a server to implement the object recognition method proposed in the embodiments of the present application. Specifically, the server 11 can obtain object sample data in a target object recognition scenario, perform data enhancement processing on the object sample data to obtain enhanced object sample data, perform feature extraction on the enhanced object sample data through multiple different object recognition models to obtain homologous object features extracted by each object recognition model, obtain a robustness measurement score corresponding to each object recognition model based on the homologous object features extracted by each object recognition model, obtain an accuracy measurement score corresponding to each object recognition model based on the reference object feature information extracted by each object recognition model for the object sample data, and select a target object recognition model that adapts to the target object recognition scenario from the multiple different object recognition models according to the accuracy measurement score and the robustness measurement score of each object recognition model, so as to perform object recognition through the target object recognition model. Then, the server 11 can deploy the target object recognition model to an application terminal 10 in the target object recognition scenario.

[0062] The following will be described in detail, and it should be noted that the order of the following embodiments is not limited as the preferred order of the embodiments.

[0063] The embodiments of the present application will be described from the perspective of an object recognition device, which can be integrated in a computer device, which can be a server or a terminal or the like.

[0064] As Figure 2 An object recognition method is provided, and the specific flow includes:

[0065] 101. Obtain object sample data in a target object recognition scenario, and perform data augmentation on the object sample data to obtain augmented object sample data.

[0066] In an embodiment, object recognition technology has been widely applied in multiple application scenarios. For example, object recognition technology can be applied to security, payment, access control, object recognition, and the like. Among them, the target object recognition scenario can refer to the scenario for which the developer deploys the model. For example, when the developer needs to deploy a target object recognition model for a security scenario, the target object recognition scenario can be a security scenario. For another example, when the developer needs to deploy a target object recognition model for an access control scenario, the target object recognition scenario can be a payment scenario.

[0067] Among them, the object proposed in the embodiments of the present application can refer to various contents. For example, the object can be a person's face, a person's palm, or a person's contour, and the like. For another example, the object can be various objects, and the like.

[0068] Among them, the object sample data can include data generated by the object in the target object recognition scenario. Among them, the object sample data can have multiple forms. For example, the object sample data can be a video or an image, and the like. For example, the object sample data can be an image. For example, when a face is recognized, the object sample data can be an image of the face of the object. For another example, when a certain object is recognized, the object sample data can be an image of the object, and the like.

[0069] In an embodiment, when object recognition is performed, object recognition model can be used for recognition.

[0070] Among them, the object recognition model can be a trained artificial intelligence model that can realize the function of object recognition.

[0071] Among them, artificial intelligence (Artificial Intelligence, AI) is to use digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0072] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software level technology. Artificial intelligence basic technology generally includes, such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning and other major directions.

[0073] Machine learning (ML) is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a specialized study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning. Among them, reinforcement learning is a field of machine learning that emphasizes how to act based on the environment to achieve maximum expected benefits. Deep reinforcement learning is a combination of deep learning and reinforcement learning, which uses deep learning techniques to solve reinforcement learning problems.

[0074] For example, the object recognition model can be a convolutional neural network (CNN), a de-convolutional network (DN), a deep neural network (DNN), a deep convolutional inverse graphics network (DCIGN), a region-based convolutional network (RCNN), a self-attentive sequential recommendation model (SASRec), a faster region-based convolutional network (Faster RCNN), a neural network model based on Wasserstein distance, or a bidirectional encoder representations from transformers (BERT) model, etc.

[0075] In an embodiment, in order to improve the efficiency of object recognition, a plurality of different object recognition models can be pre-trained, and then a target object recognition model most suitable for a target object recognition scene is selected from the plurality of object recognition models.

[0076] The plurality of different object recognition models can refer to a plurality of object recognition models with different attribute information. The attribute information can refer to information describing the attributes of the model. For example, the attribute information can include the capacity, network structure, model parameters, and objective function of the model, etc. For example, the plurality of different object recognition models can refer to a plurality of object recognition models with different attribute information such as model capacity, network structure, model parameters, and objective parameters.

[0077] In an embodiment, in the present application, the accuracy measurement score and the robustness measurement score of each object recognition model for identifying the object sample data of the target object recognition scene can be evaluated, so as to select the target object recognition model most suitable for the target object recognition scene.

[0078] In order to evaluate the accuracy measurement score and the robustness measurement score of the object recognition model, the object sample data can be subjected to data enhancement processing to obtain enhanced object sample data.

[0079] In the present application, a plurality of different data enhancement methods can be used to perform data enhancement processing on the object sample data to obtain enhanced object sample data.

[0080] Specifically, the step of "performing data enhancement processing on the object sample data to obtain enhanced object sample data" can include:

[0081] obtaining a plurality of different data enhancement methods;

[0082] determining the enhancement parameters corresponding to each data enhancement method;

[0083] based on the enhancement parameters corresponding to each data enhancement method, using a plurality of different data enhancement methods to perform data enhancement processing on the object sample data to obtain enhanced object sample data.

[0084] The data enhancement method can include random cropping, random occlusion, Gaussian blur, rotation, noise addition, edge gradient extraction, etc.

[0085] In an embodiment, each data augmentation manner has a corresponding augmentation parameter, which determines the degree of data augmentation. For example, when the data augmentation manner is rotation, the augmentation parameter can be the selected angle. For another example, when the data augmentation manner is Gaussian blur, the augmentation parameter can be the degree of Gaussian blur, and so on. Since the embodiments of the present application evaluate multiple different object recognition models, in order to ensure the effectiveness of the evaluation, the data augmentation manner and the corresponding augmentation parameter can be set to consistent hyperparameters in each evaluation model. Therefore, the augmentation parameter corresponding to each data augmentation manner can be determined, so as to ensure that the augmentation parameter corresponding to each data augmentation manner is consistent in each object recognition model.

[0086] Then, the object sample data can be processed by using multiple different data augmentation manners to obtain augmented object sample data.

[0087] In an embodiment, the augmented object sample data obtained by data augmentation on the same object sample data can be homologous object sample data, i.e., indicating that the augmented object sample data is derived from the same object sample data. For example, it is assumed that the object sample data is an image. It is assumed that the object sample image A1 is subjected to data augmentation to obtain homologous object sample images A1~AN. Wherein, the semantic identity information of the homologous object sample images A1~AN comes from the same object identity (Identity Document, ID), i.e., the object recorded in the homologous object sample images A1~AN is the same.

[0088] 102, extracting features from the augmented object sample data by using multiple different object recognition models to obtain homologous object features extracted by each object recognition model.

[0089] Wherein, the homologous object features can refer to the features obtained by extracting features from the homologous object sample data. In an embodiment, since the semantic identity information of the homologous object sample data comes from the same object identity, the homologous object features are highly similar. For example, it is assumed that the homologous object sample images A1~AN are subjected to feature extraction to obtain homologous object features feaA1~feaAN, wherein feaA1~feaAN should be very similar.

[0090] In an embodiment, before the augmented object sample data is subjected to feature extraction by using multiple different object recognition models, a to-be-trained object recognition model can be obtained, and then the to-be-trained object recognition model can be trained to obtain the object recognition model.

[0091] Specifically, as Figure 3As shown, before the step "perform feature extraction on the enhanced object sample data through a plurality of different object recognition models to obtain homologous object features extracted by each object recognition model", the following steps can be included:

[0092] obtain an object recognition model to be trained and object training sample data;

[0093] perform spatial feature extraction on the object training sample data by using the object recognition model to be trained to obtain spatial structure information corresponding to the object training sample data;

[0094] perform full-connection mapping processing on the spatial structure information corresponding to the object training sample data by using the object recognition model to be trained to obtain a recognition result corresponding to the object training sample data;

[0095] calculate model loss information based on preset label information of the object training sample data and the recognition result corresponding to the object training sample data;

[0096] adjust model parameters of the object recognition model to be trained based on the model loss information to obtain an object recognition model.

[0097] The object recognition model to be trained can include a model whose performance does not meet requirements and needs to be trained.

[0098] The object training sample data can include training data used when training the object recognition model to be trained. The object training sample data can be an image, a video, or the like.

[0099] In an embodiment, spatial feature extraction can be performed on the object training sample data by using the object recognition model to be trained to obtain spatial structure information corresponding to the object training sample data.

[0100] The spatial structure information corresponding to the object training sample data can indicate an object spatial structure of an object in the object training sample data. For example, when the object refers to a face, the object structure information can indicate an inter-brow distance of the object, a straightness of a nose bridge, a depth of an eye socket, or the like.

[0101] In an embodiment, convolution calculation, nonlinear activation function calculation, and pooling calculation can be performed on the object training sample data by using the object recognition model to be trained to obtain the spatial structure information corresponding to the object training sample data. The nonlinear activation function can include a Relu function, a Sigmoid function, or a Tanh function, or the like. The pooling calculation can include average pooling calculation or maximum pooling calculation, or the like.

[0102] In an embodiment, the spatial structure information corresponding to the object training sample data can be fully connected mapping processed by using the object recognition model to be trained, to obtain a recognition result corresponding to the object training sample data.

[0103] The recognition result corresponding to the object training sample data can be used to indicate identity information of the object in the object training sample data. For example, the recognition result of the object training sample data can indicate whether the object in the object training sample data is Zhang San or Li Si, and the like. For another example, the recognition result of the object training sample data can indicate whether the object in the object training sample data is Zhang San, and the like. For another example, when the animal is recognized, the recognition result can indicate whether the object is a cat or a dog, and the like.

[0104] In an embodiment, the model loss information can be calculated based on the preset label information of the object training sample data and the recognition result corresponding to the object training sample data. For example, the spatial structure information corresponding to the object training sample data can be input into the fully connected mapping unit in the object recognition model to be trained, to obtain a recognition result output by the fully connected mapping unit. Then, the preset label information of the object training sample data and the recognition result corresponding to the object training sample data can be input into a preset loss calculation function, to obtain the model loss information. The preset loss calculation function can include a softmax function, a softmax function of each type plus a margin, and the like.

[0105] In an embodiment, the model parameters of the object recognition model to be trained can be adjusted based on the model loss information, to obtain the object recognition model.

[0106] For example, the model parameters of the object recognition model to be trained can be adjusted by using a gradient descent method based on the model loss information, to obtain the object recognition model. The gradient descent method can include a stochastic gradient descent method, a stochastic gradient descent with a quantity term, an Adam optimization algorithm or an Adagard optimization algorithm, and the like.

[0107] The above steps are repeatedly performed until the training result meets a training termination condition. The termination condition of the model training can be that the number of iterations meets a set value, or the model loss information is less than a set value, so that the training of the model can be completed.

[0108] The object recognition model to be trained of different capacities, different network structures, and different target functions is trained by using the above process, to obtain a plurality of object recognition models.

[0109] In an embodiment, the embodiments of the present application can screen out a target object recognition model most suitable for a target object recognition scene in a plurality of different object recognition scenes. In order to screen out the target object recognition model, the object recognition model needs to be used to recognize the enhanced object sample data of the target object recognition scene to obtain homologous object features. Then, the robustness measurement score of the object recognition model can be calculated according to the homologous object features, and the degree of adaptation of the object recognition model and the target object recognition scene can be judged according to the robustness measurement score, so as to screen out the target object recognition model.

[0110] For example, taking an image as an example, assuming that the enhanced object sample data can include a plurality of homologous object images, when the robustness measurement score of the object recognition model is better, it can be indicated that the effect of the object recognition model in recognizing the plurality of homologous object images of the target object recognition scene is better, so the object recognition model and the target object recognition scene are more suitable. On the contrary, when the robustness measurement score of the object recognition model is poor, it can be indicated that the effect of the object recognition model in recognizing the plurality of homologous object images of the target object recognition scene is poor, so the object recognition model and the target object recognition scene are not suitable.

[0111] Among them, the plurality of homologous object images can include images generated from the same original object image, and the plurality of homologous object images all record the object information of the same object. For example, the plurality of homologous object images are all object images obtained by data enhancement on the same original object image.

[0112] In an embodiment, assuming that there are M object recognition models. The M object recognition models can be used to extract features from the plurality of homologous object images. Then, each object recognition model has corresponding homologous object features.

[0113] In an embodiment, when considering the degree of adaptation between the object recognition model and the target object recognition scene, the accuracy measurement score of the object recognition model can also be combined to judge the degree of adaptation between the object recognition model and the target object recognition scene. Among them, when evaluating the accuracy measurement score of the object recognition model, the accuracy measurement score of the object recognition model is judged by the recognition ability of the object recognition model on the object sample data. Therefore, the object recognition model can be used to extract features from the object sample data to obtain reference object feature information. Among them, the reference object feature information can be used to indicate the object space features of the object in the object sample data. In an embodiment, when there are a plurality of object sample data, the reference object feature information can include the object features of the objects in the plurality of object sample data. For example, the reference object feature information can be a matrix, and each row vector in the matrix can be an object feature of an object in an object sample data.

[0114] In addition, in addition to the accuracy measurement score of the object recognition model that can be measured by using the object sample data, the accuracy measurement score of the object recognition model can also be evaluated by the general recognition ability of the object recognition model for various scenes.

[0115] Therefore, the reference object feature information can be obtained by using the object recognition model to recognize a plurality of reference object data.

[0116] The plurality of reference object data can include images or videos of various object recognition scenes. For example, the plurality of reference object data can be object images sampled from an evaluation data set used to evaluate the accuracy of the object recognition model. The object images in the evaluation data set are not specifically for a certain application scene, but cover object images of many object recognition scenes.

[0117] For example, when the reference object feature information is obtained by the reference object data, the reference object feature information can be used to describe the object space features of the objects in the reference object data. In an embodiment, when there are a plurality of reference object data, the reference object feature information can include the object features of the objects in the plurality of reference object data. For example, the reference object feature information can be a matrix, and each row vector in the matrix can be an object feature of an object in a reference object data.

[0118] In an embodiment, the embodiments of the present application can include:

[0119] Obtaining a plurality of reference object data;

[0120] Using the object recognition model to perform feature extraction on the plurality of reference object data to obtain the reference object feature information.

[0121] In an embodiment, assuming that the reference object data is a reference object image, a plurality of reference object images can be obtained from a preset reference object image set. The plurality of reference object images in the preset reference object image set all carry label information with identity annotations, so that the accuracy of the object recognition model can be evaluated according to the label information of the reference object image.

[0122] Then, the object recognition model can be used to perform feature extraction on the plurality of reference object images to obtain the reference object feature information.

[0123] In an embodiment, assuming that there are M object recognition models. The M object recognition models can be used to perform feature extraction on the plurality of reference object images. Then, each object recognition model has corresponding reference object feature information.

[0124] 103、Based on the homologous object features extracted by each object recognition model, a robustness measurement score corresponding to each object recognition model is obtained.

[0125] In an embodiment, in order to determine the object recognition model that is most suitable for the target object recognition scene, the embodiments of the present application can evaluate the object recognition model from the robustness and accuracy of the model to screen out the target object recognition model.

[0126] In an embodiment, the robustness of the model can refer to the performance of the input interference or adversarial samples. In order to test the robustness of each object recognition model, the similarity of the multiple homologous object features recognized by the object recognition model can be calculated to obtain the homologous object feature similarity of the object recognition model. Then, based on the homologous object feature similarity corresponding to each object recognition model, the robustness measurement score corresponding to the object recognition model is determined.

[0127] In an embodiment, in order to obtain the robustness measurement score corresponding to each object recognition model, the similarity of the multiple homologous object features recognized by each object recognition model can be calculated to obtain the homologous object feature similarity of the object recognition model. Then, based on the homologous object feature similarity corresponding to each object recognition model, the robustness measurement score corresponding to the object recognition model is determined.

[0128] Specifically, the step of "obtaining the robustness measurement score corresponding to each object recognition model based on the homologous object features extracted by each object recognition model" can include:

[0129] combining the multiple homologous object features recognized by the object recognition model to obtain a homologous object feature group;

[0130] calculating the similarity distance between the homologous object features in the homologous object feature group;

[0131] de-noising the similarity distance of the homologous object feature group to obtain the homologous object feature similarity of the object recognition model;

[0132] determining the robustness measurement score corresponding to each object recognition model based on the homologous object feature similarity corresponding to each object recognition model.

[0133] The homologous object feature similarity can indicate the similarity between the multiple homologous object features corresponding to the object recognition model.

[0134] For example, assuming that there are M object recognition models, and each object recognition model corresponds to N homologous object features. Then, the similarity between the N homologous object features of the object recognition features can be calculated to obtain the homologous object feature similarity corresponding to the object recognition model.

[0135] In an embodiment, the similarity calculation on the plurality of homogenous object features identified by the object recognition model can refer to the similarity calculation between the plurality of homogenous object features of the object recognition model. Thus, the plurality of homogenous object features identified by the object recognition model can be combined to obtain a homogenous object feature group, and then the similarity distance between the homogenous object features of the homogenous object feature group can be calculated.

[0136] In an embodiment, the combination of the plurality of homogenous object features identified by the object recognition model can refer to the combination of the plurality of homogenous object features identified by the object recognition model two by two. For example, there are three homogenous object features, which are homogenous object feature A, homogenous object feature B and homogenous object feature C. Then, the homogenous object feature A and the homogenous object feature B can be combined to obtain a homogenous object feature group. Similarly, the homogenous object feature B and the homogenous object feature C can be combined to obtain a homogenous object feature group, and the homogenous object feature A and the homogenous object feature C can be combined to obtain a homogenous object feature group.

[0137] Then, the similarity distance between the homogenous object features of the homogenous object feature group can be calculated. For example, the similarity distance between the homogenous object feature A and the homogenous object feature B, the similarity distance between the homogenous object feature B and the homogenous object feature C, and the similarity distance between the homogenous object feature A and the homogenous object feature C can be calculated respectively.

[0138] In an embodiment, the similarity distance is used to indicate the similarity between the homogenous object features. Generally, the smaller the similarity distance, the higher the similarity between the homogenous object features. Conversely, the larger the similarity distance, the lower the similarity between the homogenous object features.

[0139] In an embodiment, there are various ways to calculate the similarity distance between the homogenous object features of the homogenous object feature group. For example, the similarity distance between the homogenous object features of the homogenous object feature group can be calculated based on the Euclidean distance, the Hamming distance or the cosine distance.

[0140] In an embodiment, after obtaining the similarity distance between the homogenous object features, due to the calculation method or calculation abnormality, the similarity distance between the homogenous object features may have abnormal values. For example, when the cosine distance is used to calculate the similarity distance between the homogenous object features of the homogenous object feature group, the similarity distance may have a negative value. For another example, when the Euclidean distance is used to calculate the similarity distance, due to the calculation abnormality, the similarity distance may be abnormal, for example, the similarity distance is very large or very small, etc. Therefore, the similarity distance of the homogenous object feature group can be denoised to improve the accuracy of the similarity of the homogenous object features of the object recognition model.

[0141] The similarity distance of the feature set of the homologous object can be denoised in various manners to obtain the similarity of the homologous object features of the object recognition model.

[0142] For example, the abnormal similarity distance can be replaced by a preset mean value. For another example, the abnormal similarity distance can be removed. For another example, the similarity distance that is a negative number can be converted into a positive value.

[0143] In an embodiment, the similarity of the homologous object features recognized by the object recognition model can be calculated according to the following formula to obtain the similarity of the homologous object features of the object recognition model, specifically:

[0144] b ij =1-cos(f i ,j)

[0145] wherein b ij may refer to the similarity of the homologous object features, and the number of the similarity of the homologous object features can be equal to the number of the feature set of the homologous object. Wherein i∈(0, N], j∈(0, N], ≠j, N can refer to the number of the homologous object features recognized by the object recognition model. f i and f j may refer to the homologous object features.

[0146] In an embodiment, the similarity of the homologous object features of each object recognition model can be obtained through the above steps. For example, there are M object recognition models B1-BM. The similarity of the homologous object features of the object recognition models B1-BM can be obtained through the above steps.

[0147] In an embodiment, after obtaining the similarity of the homologous object features of each object recognition model, the robustness measurement score corresponding to the object recognition model can be determined for each object recognition model.

[0148] Wherein the robustness measurement score can be used to indicate the robustness of the object recognition model. For example, through the robustness measurement score, the performance of the object recognition model in resisting interference or adversarial samples in the target object recognition scene can be known.

[0149] In an embodiment, after obtaining the similarity of the homologous object features of the object recognition model, the similarity of the homologous object features can be further processed to obtain the robustness measurement score.

[0150] Specifically, the step of “determining, for each object recognition model, the robustness measurement score corresponding to the object recognition model based on the similarity of the homologous object features corresponding to the object recognition model” can include:

[0151] The de-noised similarity distance of each homologous object feature group in the homologous object feature similarity corresponding to the object recognition model is fused to obtain a fused similarity distance.

[0152] The fused similarity distance is normalized to obtain a robustness measurement score corresponding to the object recognition model.

[0153] In an embodiment, the homologous object feature similarity corresponding to the object recognition model includes de-noised similarity distances of multiple homologous object feature groups. For example, the homologous object feature similarity corresponding to the object recognition model can include a de-noised similarity distance between homologous object feature A and homologous object feature B, a de-noised similarity distance between homologous object feature B and homologous object feature C, and a de-noised similarity distance between homologous object feature A and homologous object feature C. Therefore, the de-noised similarity distance of each homologous object feature group in the homologous object feature similarity corresponding to the object recognition model can be fused to obtain a fused similarity distance. For example, the de-noised similarity distance of each homologous object feature group can be added to obtain the fused similarity distance. For another example, the de-noised similarity distance of each homologous object feature group can be multiplied to obtain the fused similarity distance, and so on.

[0154] Then, in order to improve the efficiency of processing, the fused similarity distance can be normalized to obtain a robustness measurement score corresponding to the object recognition model. Wherein, the way of normalizing the fused similarity distance can be determined according to the fusion manner. For example, if the fused similarity distance is obtained by adding, the fused similarity distance can be divided by the number of homologous object features to obtain the robustness measurement score. For another example, if the fused similarity distance is obtained by multiplying, the fused similarity distance can be square-rooted to obtain the robustness measurement score corresponding to the object recognition model.

[0155] In an embodiment, as Figure 4As shown, assuming that the object sample data is a face image, the object recognition model can be a face recognition model. Assuming that data augmentation is performed on the face image, N homologous face images A1-AN are obtained. Among the N homologous face images, the initial face image A1 and its augmented forms A2-AN are included. The semantic identity information of the homologous face images A1-AN all comes from the same object identity (Identity Document, ID), that is, the object recorded in the homologous face images A1-AN is the same. If the face recognition model has strong robustness, the homologous face features extracted by the face recognition model from the homologous face images should be very similar, that is, the homologous face features feaA1-feaAN of the homologous face images A1-AN should be very similar. In the embodiment of the present application, the robustness measurement score of the face recognition model can be measured by the following formula:

[0156]

[0157] Wherein, N can refer to the number of homologous face features identified by the face recognition model. i And f j may refer to the homologous face feature. L1 can refer to the robustness measurement score of the face recognition model.

[0158] In an embodiment, the better the robustness of the face recognition model to the target face recognition scene data, the lower the robustness measurement score L1 of the face recognition model.

[0159] In an embodiment, by the above steps, the robustness measurement score corresponding to each face recognition model can be obtained. For example, there are M face recognition models B1-BM. By the above steps, the robustness measurement scores of the face recognition models B1-BM can be obtained.

[0160] 104、Based on the reference object feature information extracted by each object recognition model for the object sample data, an accuracy measurement score corresponding to each object recognition model is obtained.

[0161] In an embodiment, in order to determine the object recognition model most suitable for the target object recognition scene, in addition to evaluating the object recognition model from the robustness, the object recognition model can also be evaluated from the accuracy to filter out the target object recognition model.

[0162] In an embodiment, the accuracy measurement score corresponding to each object recognition model can be obtained based on the reference object feature information extracted by each object recognition model for the object sample data. Wherein, the reference object feature information extracted by the object recognition model can be identified to obtain the accuracy reference threshold corresponding to the object recognition model. Then, the accuracy measurement score corresponding to the object recognition model is calculated according to the accuracy reference threshold corresponding to the object recognition model.

[0163] Specifically, the step of "obtaining the accuracy measurement score corresponding to each object recognition model based on the reference object feature information extracted by each object recognition model for the object sample data" can include:

[0164] identifying the reference object feature information extracted by each object recognition model to obtain the accuracy reference threshold corresponding to each object recognition model;

[0165] According to the label information corresponding to the object sample data, the recall threshold list corresponding to each object recognition model is identified in the reference object feature information of each object recognition model;

[0166] According to the accuracy reference threshold corresponding to each object recognition model, the recall threshold list corresponding to each object recognition model is discriminated to obtain the accuracy measurement score corresponding to each object recognition model.

[0167] Wherein, the accuracy reference threshold can be used to discriminate the accuracy of the object recognition model. Wherein, the accuracy reference threshold is generated according to the reference object feature information, so that the accuracy measurement score corresponding to the object recognition model is calculated through the accuracy reference threshold, which can objectively and comprehensively indicate the accuracy of the object recognition model.

[0168] In an embodiment, the reference object feature information includes object space features corresponding to a plurality of object sample data, so that the object space feature similarity between the object sample data can be calculated according to the reference object feature information. Then, the accuracy reference threshold of the object recognition model can be evaluated according to the object space feature similarity.

[0169] Specifically, the step of "identifying the reference object feature information extracted by each object recognition model to obtain the accuracy reference threshold corresponding to each object recognition model" can include:

[0170] The object sample data is combined to obtain an object sample data group;

[0171] Based on the reference object feature information, the object space feature similarity between the object sample data in the object sample data group is calculated;

[0172] According to the label information of the object sample data in the object sample data set, a set of heterogeneous image groups is identified in the object data sample set;

[0173] Based on the object space feature similarity of the object sample data set in the set of heterogeneous image groups, an accuracy reference threshold corresponding to the object recognition model is generated.

[0174] In an embodiment, due to multiple object sample data, in order to calculate the object space feature similarity between the object sample data, the multiple object sample data can be combined to obtain an object sample data set. For example, the multiple object sample data can be combined two by two to obtain multiple object sample data sets. For example, assuming that the object sample data is an image, there are object image A, object image B and object image C. Then, the object image A, object image B and object image C can be combined two by two to obtain 3 object image groups.

[0175] In an embodiment, since the reference object feature information can include object features corresponding to multiple object sample data, the object space feature similarity between the object sample data in the object sample data set can be calculated based on the reference object feature information. For example, the object features corresponding to the object sample data in the object sample data set can be extracted from the reference object feature information, and then the object space feature similarity between the object sample data is calculated according to the object features. There are various methods to calculate the object space feature similarity. For example, the object features corresponding to the reference object images can be calculated by cosine or multiplied, etc., to obtain the object space feature similarity between the reference object images

[0176] In an embodiment, a set of heterogeneous image groups can be identified in the object data sample set according to the label information of the object sample data in the object sample data set.

[0177] The label information corresponding to the object sample data can be used to indicate the identity of the object in the object sample data. For example, the label information of the object sample data can indicate whether the object in the object sample data is Zhang San or Li Si. For another example, the label information of the object sample data can indicate whether the object in the object sample data is Zhang San, etc.

[0178] The object sample data set in the heterogeneous image group set includes object sample data with inconsistent label information. For example, object sample data A and object sample data B are combined to obtain an object sample data set. The label information of object sample data A and object sample data B is inconsistent, and the object sample data set can be added to the heterogeneous image group set. For another example, object sample data A and object sample data C are combined to obtain an object sample data set. The label information of object sample data A and object sample data C is consistent, and the object sample data set will not be added to the heterogeneous image group set.

[0179] In an embodiment, the accuracy reference threshold of the object recognition model can be generated based on the object space feature similarity corresponding to the object image group in the heterogeneous image group set.

[0180] Specifically, the step of "generating the accuracy reference threshold of the object recognition model based on the object space feature similarity corresponding to the object image group in the heterogeneous image group set" can include:

[0181] The object space feature similarity between the reference object image groups in the heterogeneous image group set is sorted to obtain a sorted heterogeneous image group set.

[0182] The false positive threshold list is screened out from the sorted heterogeneous image group set.

[0183] The accuracy reference threshold of the object recognition model is determined from the object space feature similarity in the false positive threshold list according to the preset false positive rate.

[0184] In an embodiment, the object space feature similarity between the object sample data sets in the heterogeneous image group set can be sorted to obtain a sorted heterogeneous image group set.

[0185] For example, the heterogeneous image group set includes four object sample data sets, and then the object space feature similarity corresponding to the object sample data set can be sorted to obtain a sorted heterogeneous image group set. That is, the object space feature similarity corresponding to the object sample data set in the sorted heterogeneous image group set is in order, rather than chaotic.

[0186] In an embodiment, the false positive threshold list can be screened out from the sorted heterogeneous image group set. The false positive threshold list includes the object space feature similarity of the object sample data with incorrect identification. For example, it is assumed that the objects in object sample data i and object sample data j are not the same person, but the object recognition model identifies the two images as the same person, so the object space feature similarity between object sample data i and object sample data j is added to the false positive threshold list.

[0187] In an embodiment, the accuracy reference threshold corresponding to the object recognition model can be determined in the object space feature similarity of the false positive threshold list according to a preset false positive rate.

[0188] The preset false positive rate can be a preset value. The preset false positive rate can be a reference value for determining the accuracy reference threshold. The preset false positive rate is generally an empirical value, and is generally a value less than 1.

[0189] In an embodiment, the number of object space feature similarities in the false positive threshold list can be identified. Then, the number of object space feature similarities is multiplied by the preset false positive rate to obtain the position of the accuracy reference threshold in the false positive threshold list. Then, the object space feature similarity corresponding to the position is determined as the accuracy reference threshold. For example, assuming that the list length of the false positive threshold list is m, that is, the false positive threshold list includes m object space feature similarities. The preset false positive rate is k. According to the preset false positive rate k, the k*mth object space feature similarity in the false positive threshold list is taken as the accuracy reference threshold.

[0190] In an embodiment, through the above steps, the accuracy reference threshold corresponding to each object recognition model can be identified. For example, there are M object recognition models B1-BM. Through the above steps, the accuracy reference threshold corresponding to the object recognition models B1-BM can be obtained.

[0191] In an embodiment, after obtaining the accuracy reference threshold, for each object recognition model, the accuracy measurement score corresponding to the object recognition model is calculated according to the accuracy reference threshold corresponding to the object recognition model. For example, according to the label information corresponding to the object sample data, the recall threshold list corresponding to each object recognition model can be identified in the reference object feature information of each object recognition model. Then, according to the accuracy reference threshold corresponding to each object recognition model, the recall threshold list corresponding to each object recognition model is discriminated to obtain the accuracy measurement score corresponding to each object recognition model.

[0192] The step of "identifying the recall threshold list corresponding to each object recognition model in the reference object feature information of each object recognition model according to the label information corresponding to the object sample data" can include:

[0193] Identifying a set of homologous image groups in the object sample data set based on the label information corresponding to the object sample data;

[0194] Identifying the recall threshold list in the set of homologous image groups;

[0195] In an embodiment, the same-source image group set can be identified in the object sample data set based on the label information corresponding to the object sample data.

[0196] The same-source image group set includes the object sample data set with consistent label information. For example, the object sample data A and the object sample data C are combined to obtain the object sample data set. The label information of the object sample data A and the object sample data C is consistent, and the reference object image group can be added to the different-source image group set.

[0197] Then, the recall threshold list can be identified in the same-source image group set. The object space feature similarity between the object sample data sets in the different-source image group set can be sorted to obtain the sorted different-source image group set. Then, the false positive threshold list is screened out in the sorted different-source image group set.

[0198] For example, the same-source image group set includes four reference object image groups. Then, the object space feature similarity corresponding to the reference object image group can be sorted to obtain the sorted same-source image group set. That is, the object space feature similarity corresponding to the object sample data set in the sorted same-source image group set is ordered, rather than chaotic.

[0199] Then, the recall threshold list can be screened out in the sorted different-source image group set. The object space feature similarity between the object sample data sets with correct identity recognition is included in the recall threshold list. For example, it is assumed that the objects in the object sample data i and the object sample data j are the same person, and the object recognition model identifies that the two sample data are the same person. Therefore, the object space feature similarity between the object sample data i and the object sample data j is added to the recall threshold list.

[0200] In an embodiment, the accuracy reference threshold corresponding to each object recognition model can be used to distinguish the recall threshold list corresponding to each object recognition model to obtain the accuracy measurement score corresponding to each object recognition model.

[0201] Specifically, the step of “distinguishing the recall threshold list corresponding to each object recognition model according to the accuracy reference threshold corresponding to each object recognition model to obtain the accuracy measurement score corresponding to each object recognition model” can include:

[0202] Comparing the object space feature similarity in the recall threshold list with the accuracy reference threshold to obtain a comparison result;

[0203] Determining the misjudgment feature similarity in the recall threshold list according to the comparison result;

[0204] The number of misjudgment feature similarities is compared with the number of object space feature similarities in the recall threshold list to obtain an accuracy measurement score corresponding to the object recognition model.

[0205] In an embodiment, the object space feature similarities in the recall threshold list can be compared with an accuracy reference threshold. For example, it can be compared which object space feature similarities in the recall threshold list are less than the accuracy reference threshold. Then, the misjudgment feature similarities can be determined in the recall threshold list according to the comparison result. The misjudgment feature similarities can refer to the object space feature similarities in the recall threshold list that are wrongly judged. For example, the object sample data set includes base object sample data with consistent label information. If the calculated object space feature similarity corresponding to the object sample data set is less than the accuracy reference threshold, it can be indicated that the object space feature similarity is a misjudgment feature similarity, so the object recognition model can identify the object sample data with consistent label information as inconsistent.

[0206] Generally, the better the accuracy performance of the object recognition model is, the fewer the misjudgment feature similarities corresponding to the object recognition model are. Conversely, the worse the accuracy performance of the object recognition model is, the more the misjudgment feature similarities corresponding to the object recognition model are. Therefore, the misjudgment feature similarities can be determined in the recall threshold list according to the comparison result. Then, the accuracy measurement score corresponding to the object recognition model can be calculated according to the number of misjudgment feature similarities and the number of object space feature similarities in the recall threshold list.

[0207] For example, the number z of similarities less than the accuracy reference threshold t can be obtained according to the accuracy reference threshold t, and the accuracy measurement score is The accuracy measurement score can refer to the false negative rate, so generally, the lower the accuracy measurement score is, the better the recognition accuracy of the object recognition model is.

[0208] In an embodiment, it is assumed that the reference object feature information is a matrix, and the matrix is represented as F W×N The matrix has a dimension of WxN, where N is the number of reference object images, and W is the dimension of the reference object feature information. Then, the object space feature similarity between the plurality of object sample data can be obtained by calculating F T F to obtain a similarity matrix. The dimension of the similarity matrix is NxN, and the lower triangular matrix of the similarity matrix is taken, and the column value of the lower triangular matrix is the object space feature similarity between the object sample data i and the object sample data j.

[0209] Then, whether the objects in the object sample data i and the object sample data j are the same identity can be obtained through the label information corresponding to each object sample data. Then, the object space feature similarities between the object sample data with different identities in the lower triangular matrix can be taken out through the label information corresponding to each object sample data, to obtain a first similarity set. Then, the object space feature similarities in the first similarity set can be sorted from small to large to obtain a sorted similarity set. The object space feature similarities that are incorrectly recognized are taken out from the sorted similarity set to form a false positive threshold list S e , and it is assumed that the length of the false positive threshold list S e is m.

[0210] That is, the object space feature similarities between the object sample data with incorrect identity recognition are included in the false positive threshold list S e . For example, it is assumed that the objects in the object sample data i and the object sample data j are not the same person, but the object recognition model identifies the two images as the same person, so the object space feature similarity between the object sample data i and the object sample data j is added to the false positive threshold list S e .

[0211] In addition, the object space feature similarities between the object sample data with the same identity in the lower triangular matrix can also be taken out through the label information corresponding to each object sample data to obtain a second similarity set. Then, the object space feature similarities in the second similarity set can be sorted from small to large to obtain a sorted similarity set. The object space feature similarities that are correctly recognized are taken out from the sorted similarity set to form a recall threshold list S a , and it is assumed that the length of the recall threshold list S a is l.

[0212] That is, the object space feature similarities between the object sample data with correct identity recognition are included in the false positive threshold list S α . For example, it is assumed that the objects in the object sample data i and the object sample data j are the same person, and the object recognition model identifies that the two images are the same person, so the object space feature similarity between the reference object image i and the reference object image j is added to the recall threshold list S a .

[0213] In an embodiment, the model false negative rate under a preset false positive rate can be tested, and the lower the false negative rate under the same false positive rate, the more accurate the model. The preset false positive rate can be set in advance as k. According to the preset false positive rate k, the false positive threshold list S eThe kth similarity of the mth object space feature is taken as an accuracy reference threshold t. Then, a recall threshold list S can be obtained according to the accuracy reference threshold t a If the number of similarities less than the accuracy reference threshold t is z, then the accuracy measurement score is The accuracy measurement score can refer to a false alarm rate. In general, the lower the accuracy measurement score, the better the recognition accuracy of the object recognition model.

[0214] In an embodiment, the accuracy measurement score of each object recognition model can be obtained through the above steps. For example, there are M object recognition models B1 to BM. Through the above steps, the accuracy measurement scores of the object recognition models B1 to BM can be obtained.

[0215] 105. Select a target object recognition model that adapts to the target object recognition scene from the plurality of different object recognition models according to the accuracy measurement score and the robustness measurement score of each object recognition model, to perform object recognition through the target object recognition model.

[0216] In an embodiment, a target object recognition model that adapts to the target object recognition scene can be selected from the plurality of different object recognition models according to the accuracy measurement score and the robustness measurement score of each object recognition model.

[0217] Specifically, the step of "selecting a target object recognition model that adapts to the target object recognition scene from the plurality of different object recognition models according to the accuracy measurement score and the robustness measurement score of each object recognition model" can include:

[0218] Fusing the accuracy measurement score and the robustness measurement score of each object recognition model to obtain matching degree information of each object recognition model;

[0219] Sorting the matching degree information of the object recognition model to obtain sorted matching degree information of the object recognition model;

[0220] Selecting a target object recognition model that adapts to the target object recognition scene from the plurality of object recognition models based on the sorted matching degree information.

[0221] The matching degree information can indicate the matching degree between the object recognition model and the data in the target object recognition scene.

[0222] In an embodiment, the accuracy measurement score and the robustness measurement score of each object recognition model can be fused according to the following formula to obtain the matching degree information of each object recognition model:

[0223] L = aL1 + bL2

[0224] Here, α and β can refer to the fusion weights, which are generally preset. For example, the fusion weight α can be 0.2 and the fusion weight β can be 0.8.

[0225] In one embodiment, the matching degree information of object recognition models can be sorted to obtain sorted matching degree information. For example, the matching degree information of object recognition models can be sorted from smallest to largest to obtain sorted matching degree information. Then, based on the sorted matching degree information, a target object recognition model suitable for the target object recognition scenario can be selected from multiple object recognition models. For example, since the matching degree information is obtained by fusing accuracy and robustness scores, the smaller the matching degree information, the better the match between the object recognition model and the data in the target object recognition scenario. Therefore, the object recognition model with the smallest matching degree information can be selected as the target object recognition model.

[0226] In one embodiment, after selecting the target object recognition model, the target object recognition model can be deployed to the application terminal of the target object recognition scenario, so that the application terminal can use the target object recognition model to perform object recognition processing.

[0227] In one embodiment, such as Figure 5 The diagram illustrates the overall flow of the proposed method, which mainly comprises two stages: a training and evaluation stage, and a deployment stage. In the training and evaluation stage, multiple different object recognition models are first trained. Then, the robustness and accuracy of the obtained object recognition models in the target object recognition scenario are evaluated. Robustness to the target object recognition scenario is primarily achieved using a self-supervised contrastive learning scheme. Accuracy is evaluated by assessing the model's recall rate under a fixed false positive rate. Finally, considering both the robustness and accuracy of the model in relation to the target object recognition scenario, the target object recognition model that best matches the scenario is selected. In the deployment stage, only the target object recognition model needs to be exported and deployed. This method does not require modification to the original object recognition network to obtain an object recognition model that matches the target object recognition scenario.

[0228] In the embodiments of the present application, the object sample data in the target object recognition scene and the enhanced object sample data corresponding to the object sample data can be obtained; the enhanced object sample data is subjected to feature extraction by a plurality of different object recognition models to obtain homologous object features; based on the homologous object features, a robustness measurement score of each object recognition model is obtained; based on the reference object feature information of each object recognition model, an accuracy measurement score corresponding to each object recognition model is obtained; according to the accuracy measurement score and the robustness measurement score of each object recognition model, a target object recognition model suitable for the target object recognition scene can be accurately selected from the plurality of object recognition models to perform object recognition by the target object recognition model. The embodiments of the present application comprehensively evaluate the adaptation degree between the object recognition model and the target object recognition scene by two indicators, which can more comprehensively indicate the adaptation degree between the object recognition model and the target object recognition scene, so that the target object recognition model most suitable for the target object recognition scene can be accurately selected from the plurality of object recognition models.

[0229] According to the method described in the above embodiments, the following will be further described by way of example.

[0230] The object recognition method proposed in the embodiments of the present application will be described taking the object as a face as an example, and the face recognition method integrated in the server will be taken as an example to introduce the method of the embodiments of the present application.

[0231] In an embodiment, as shown in Figure 6 An object recognition method, the specific process is as follows:

[0232] 201. The server obtains a plurality of different face recognition models, homologous face features extracted by the face recognition models in a target face recognition scene, and reference face feature information extracted by the face recognition models.

[0233] In an embodiment, before obtaining a plurality of different face recognition models, the to-be-trained face recognition model can be trained to obtain the face recognition model.

[0234] For example, the to-be-trained face recognition model can be used to perform convolution calculation, nonlinear activation function calculation and pooling calculation on the face image sample to obtain spatial structure information corresponding to the face image sample. The nonlinear activation function can include a Relu function, a Sigmoid function or a Tanh function, etc. The pooling calculation can include average pooling calculation or maximum pooling calculation, etc.

[0235] Then, the spatial structure information corresponding to the face image sample can be input into a full connection mapping unit in the face recognition model to be trained, to obtain a recognition result output by the full connection mapping unit. Then, the preset label information of the face image sample and the recognition result corresponding to the face image sample can be input into a preset loss calculation function, to obtain model loss information. The preset loss calculation function can include a softmax function, a margin type softmax function of each class, and the like.

[0236] By continuously repeating the above steps, until the training result meets the training termination condition. The condition for terminating the model training is generally set to meet the set value of the number of iterations, or the model loss information is less than the set value, so that the model training can be completed.

[0237] By the above process, different capacity, different network structure, different target function of the face recognition model to be trained are trained, to obtain a plurality of face recognition models.

[0238] In an embodiment, in order to evaluate the robustness of the face recognition model, the initial face image can be subjected to data enhancement processing based on a preset enhancement parameter, to obtain a plurality of homologous face images.

[0239] There are various ways to perform data enhancement processing on the initial face image to obtain a plurality of homologous face images. For example, the initial face image can be subjected to random cropping, random occlusion, Gaussian blur, rotation, noise addition, edge gradient extraction, and the like, to obtain a plurality of homologous face images.

[0240] Then, the face recognition model can be used to extract spatial features from the plurality of homologous face images, to obtain a plurality of homologous face image spatial features. For example, the face recognition model can be used to respectively perform convolution calculation, nonlinear activation function calculation and pooling calculation on the plurality of homologous face images, to obtain a plurality of homologous face spatial features.

[0241] In an embodiment, a plurality of reference face images can be obtained from a preset reference face image set. The plurality of reference face images in the preset reference face image set are all labeled with identity label information, so that the accuracy of the face recognition model can be evaluated according to the label information of the reference face image.

[0242] Then, the face recognition model can be used to extract features from the plurality of reference face images, to obtain the reference face feature information.

[0243] 202、The server performs similarity calculation on the plurality of homologous face features recognized by the face recognition model for each face recognition model, to obtain homologous face feature similarity of the face recognition model.

[0244] 203. For each facial recognition model, the server determines the robustness score of the facial recognition model based on the similarity of the corresponding homologous facial features.

[0245] In one embodiment, such as Figure 4 As shown, assuming data augmentation is performed on the initial facial image, N homologous facial images A1 to AN are obtained. These N homologous facial images include the initial facial image A1 and its augmented forms A2 to AN. The semantic identity information of the homologous facial images A1 to AN all comes from the same object identifier (Identity Document, ID), meaning the objects recorded in the homologous facial images A1 to AN are the same. If the facial recognition model has strong robustness, the homologous facial features extracted by the facial recognition model from the homologous facial images should be very similar, that is, the homologous facial features between the homologous facial images A1 to AN should be very similar. In this embodiment, the robustness score of the facial recognition model can be measured by the following formula:

[0246]

[0247] Where N can refer to the number of homologous facial features identified by the facial recognition model. i and f j It can refer to homologous facial features. L1 can refer to the robustness score of a facial recognition model.

[0248] 204. For each facial recognition model, the server identifies the baseline facial feature information extracted by the facial recognition model to obtain the accuracy reference threshold corresponding to the facial recognition model.

[0249] 205. For each facial recognition model, the server calculates the accuracy score corresponding to the facial recognition model based on the accuracy reference threshold corresponding to the facial recognition model.

[0250] In one embodiment, it is assumed that the reference facial feature information is a matrix, and this matrix is ​​represented as F. W×N The matrix has dimensions W×N, where N is the number of reference facial images and W is the dimension of the reference facial feature information. Then, F can be calculated... T F yields a similarity matrix that describes the similarity of facial spatial features between multiple reference facial images. This similarity matrix has dimensions N×N. Taking the lower triangular matrix of this similarity matrix, the column values ​​of the lower triangular matrix represent the facial spatial feature similarity between reference facial image i and reference facial image j.

[0251] Then, whether the subjects in the reference face image i and the reference face image j are the same identity can be obtained through the label information corresponding to each reference face image. Then, the face space feature similarities between the reference face images with different identities in the lower triangular matrix can be taken out through the label information corresponding to each reference face image to obtain a first similarity set. Then, the face space feature similarities in the first similarity set can be sorted from small to large to obtain a sorted similarity set. The face space feature similarities with recognition errors can be taken out from the sorted similarity set to form a false positive threshold list S e , and it is assumed that the length of the false positive threshold list S e is m.

[0252] That is, the face space feature similarities between the reference face images with identity recognition errors are included in the false positive threshold list S e . For example, it is assumed that the subjects in the reference face image i and the reference face image j are not the same person, but the face recognition model identifies the two images as the same person, and thus the face space feature similarity between the reference face image i and the reference face image j is added to the false positive threshold list S e .

[0253] In addition, the face space feature similarities between the reference face images with the same identity in the lower triangular matrix can also be taken out through the label information corresponding to each reference face image to obtain a second similarity set. Then, the face space feature similarities in the second similarity set can be sorted from small to large to obtain a sorted similarity set. The face space feature similarities with correct recognition can be taken out from the sorted similarity set to form a recall threshold list S a , and it is assumed that the length of the recall threshold list S a is l.

[0254] That is, the face space feature similarities between the reference face images with correct identity recognition are included in the false positive threshold list S a . For example, it is assumed that the subjects in the reference face image i and the reference face image j are the same person, and the face recognition model identifies the two images as the same person, and thus the face space feature similarity between the reference face image i and the reference face image j is added to the recall threshold list S a .

[0255] In an embodiment, the model false negative rate under a preset false positive rate can be tested, and the lower the false negative rate under the same false positive rate, the more accurate the model. The preset false positive rate can be set as k in advance. According to the preset false positive rate k, the false positive threshold list S eThe kth similarity of the mth face space feature is taken as an accuracy reference threshold t. Then, a recall threshold list S can be obtained according to the accuracy reference threshold t a The accuracy measurement score is calculated according to the number z of similarities less than the accuracy reference threshold t The accuracy measurement score can refer to a false alarm rate. In general, the lower the accuracy measurement score, the better the recognition accuracy of the face recognition model.

[0256] In an embodiment, the accuracy measurement score corresponding to each face recognition model can be obtained through the above steps. For example, there are M face recognition models B1-BM. Through the above steps, the accuracy measurement scores of the face recognition models B1-BM can be obtained.

[0257] 206. The server screens a target face recognition model that adapts to the target face recognition scene from the plurality of different face recognition models according to the accuracy measurement score and the robustness measurement score of each model.

[0258] In the embodiments of the present application, the server obtains a plurality of different face recognition models, homologous face features extracted by the face recognition models in the target face recognition scene, and reference face feature information extracted by the face recognition models; the server calculates the similarity of the homologous face features recognized by the face recognition model for each face recognition model to obtain the homologous face feature similarity of the face recognition model; the server determines the robustness measurement score corresponding to the face recognition model based on the homologous face feature similarity corresponding to the face recognition model for each face recognition model; the server identifies the reference face feature information extracted by the face recognition model to obtain the accuracy reference threshold corresponding to the face recognition model for each face recognition model; the server calculates the accuracy measurement score corresponding to the face recognition model according to the accuracy reference threshold corresponding to the face recognition model for each face recognition model; and the server screens a target face recognition model that adapts to the target face recognition scene from the plurality of different face recognition models according to the accuracy measurement score and the robustness measurement score of each model. The embodiments of the present application comprehensively evaluate the adaptation degree between the face recognition model and the target face recognition scene by two indexes, which can more comprehensively indicate the adaptation degree between the face recognition model and the target face recognition scene, so that the target face recognition model that best adapts to the target face recognition scene can be accurately screened from the plurality of face recognition models.

[0259] In order to better implement the object recognition method provided in the embodiments of the present application, an object recognition device is further provided in an embodiment, which can be integrated in a computer device. The meanings of the terms are the same as those in the above object recognition method, and the specific implementation details can be referred to the description in the method embodiments.

[0260] In an embodiment, an object recognition apparatus is provided, which can be integrated in a computer device, as shown in Figure 7 The object recognition apparatus comprises a data augmentation unit 301, a feature extraction unit 302, a first score acquisition unit 303, a second score acquisition unit 304 and a screening unit 305, specifically as follows:

[0261] The data augmentation unit 301 is configured to obtain object sample data under a target object recognition scenario, and perform data augmentation processing on the object sample data to obtain augmented object sample data.

[0262] The feature extraction unit 302 is configured to perform feature extraction on the augmented object sample data by using a plurality of different object recognition models to obtain homologous object features extracted by each object recognition model.

[0263] The first score acquisition unit 303 is configured to acquire a robustness measurement score corresponding to each object recognition model based on the homologous object features extracted by each object recognition model.

[0264] The second score acquisition unit 304 is configured to acquire an accuracy measurement score corresponding to each object recognition model based on reference object feature information extracted by each object recognition model for the object sample data.

[0265] The screening unit 305 is configured to screen a target object recognition model that is adapted to the target object recognition scenario from the plurality of different object recognition models according to the accuracy measurement score and the robustness measurement score of each object recognition model, so as to perform object recognition by using the target object recognition model.

[0266] In an embodiment, the second score acquisition unit 304 can comprise:

[0267] A first identification sub-unit is configured to identify the reference object feature information extracted by each object recognition model to obtain an accuracy reference threshold corresponding to each object recognition model.

[0268] A second identification sub-unit is configured to identify a recall threshold list corresponding to each object recognition model in the reference object feature information of each object recognition model according to label information corresponding to the object sample data.

[0269] A first discrimination sub-unit is configured to discriminate the recall threshold list corresponding to each object recognition model according to the accuracy reference threshold corresponding to each object recognition model to obtain the accuracy measurement score corresponding to each object recognition model.

[0270] In an embodiment, the first identification sub-unit can comprise:

[0271] a combination module, configured to combine the object sample data to obtain an object sample data group;

[0272] a calculation module, configured to calculate object space feature similarity between the object sample data in the object sample data group based on the reference object feature information;

[0273] a recognition module, configured to recognize a set of heterogeneous image groups from the object sample data group according to label information of the object sample data in the object sample data group;

[0274] a generation module, configured to generate an accuracy reference threshold corresponding to the object recognition model based on object space feature similarity corresponding to the object sample data group in the set of heterogeneous image groups.

[0275] In an embodiment, the discrimination subunit can include:

[0276] a comparison module, configured to compare the object space feature similarity in the recall threshold list with the accuracy reference threshold to obtain a comparison result;

[0277] a determination module, configured to determine a misjudgment feature similarity from the recall threshold list according to the comparison result;

[0278] a comparison operation module, configured to perform a comparison operation on the number of misjudgment feature similarities and the number of object space feature similarities in the recall threshold list to obtain an accuracy measurement score corresponding to the object recognition model.

[0279] In an embodiment, the first score acquisition unit 303 can include:

[0280] a combination subunit, configured to combine a plurality of homologous object features recognized by the object recognition model to obtain a homologous object feature group;

[0281] a calculation subunit, configured to calculate similarity distance between the homologous object features in the homologous object feature group;

[0282] a denoising subunit, configured to perform denoising processing on the similarity distance of the homologous object feature group to obtain homologous object feature similarity of the object recognition model;

[0283] a second discrimination subunit, configured to discriminate a robustness measurement score corresponding to each object recognition model based on the homologous object feature similarity corresponding to each object recognition model.

[0284] In an embodiment, the data enhancement unit 301 can include:

[0285] The acquisition subunit is configured to acquire a plurality of different data enhancement manners.

[0286] The parameter determination subunit is configured to determine an enhancement parameter corresponding to each data enhancement manner.

[0287] The data enhancement subunit is configured to perform data enhancement processing on the object sample data based on the enhancement parameter corresponding to each data enhancement manner, to obtain enhanced object sample data.

[0288] In an embodiment, the object recognition apparatus can further include:

[0289] The acquisition unit is configured to acquire an object recognition model to be trained and object training sample data.

[0290] The feature extraction unit is configured to perform spatial feature extraction on the object training sample data by using the object recognition model to be trained, to obtain spatial structure information corresponding to the object training sample data.

[0291] The fully connected mapping unit is configured to perform fully connected mapping processing on the spatial structure information corresponding to the object training sample data by using the object recognition model to be trained, to obtain a recognition result corresponding to the object training sample data.

[0292] The calculation unit is configured to calculate model loss information based on preset label information of the object training sample data and the recognition result corresponding to the object training sample data.

[0293] The adjustment unit is configured to adjust model parameters of the object recognition model to be trained based on the model loss information, to obtain an object recognition model.

[0294] In implementation, each of the above units can be implemented as an independent entity, or can be combined as the same or several entities, and the implementation of each of the above units can refer to the method embodiments above, which will not be described herein.

[0295] The object recognition apparatus can accurately select a target object recognition model most suitable for a target object recognition scenario from a plurality of object recognition models.

[0296] Embodiments of the present application also provide a computer device, which can include a terminal or a server, for example, the computer device can be an object recognition terminal, which can be a mobile phone, a tablet computer, etc.; for another example, the computer device can be a server, such as an object recognition server, etc. As shown in FIG. 13, it shows a structural schematic diagram of a terminal related to embodiments of the present application, in particular: Figure 8

[0297] ​The computer device can include a processor 401 having one or more processing cores, a memory 402 having one or more computer readable storage media, a power supply 403, and an input unit 404, etc. Those skilled in the art can understand that Figure 8 The computer device structure shown in the figure is not a limitation on the computer device, and can include more or fewer components than shown, or combine certain components, or different component arrangements. Among them:

[0298] The processor 401 is the control center of the computer device, and connects various parts of the computer device through various interfaces and lines, and performs various functions and processes data of the computer device by running or executing software programs and / or modules stored in the memory 402, and calling data stored in the memory 402. Optionally, the processor 401 can include one or more processing cores; preferably, the processor 401 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user pages and application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 401.

[0299] The memory 402 can be used to store software programs and modules, and the processor 401 executes various functions and data processing by running the software programs and modules stored in the memory 402. The memory 402 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc.; the data storage area can store data created according to the use of the computer device, etc. In addition, the memory 402 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 402 can also include a memory controller to provide access for the processor 401 to the memory 402.

[0300] The computer device also includes a power supply 403 for powering various components, and preferably the power supply 403 can be logically connected to the processor 401 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 403 can also include one or more direct current or alternating current power supplies, a recharging system, a power supply failure detection circuit, a power supply converter or inverter, a power supply state indicator, etc. Any component.

[0301] The computer device can also include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0302] Although not shown, the computer device can further include a display unit and the like, which will not be described here. Specifically in the present embodiment, the processor 401 in the computer device will load the executable file corresponding to the process of one or more application programs into the memory 402 according to the following instructions, and run the application program stored in the memory 402 by the processor 401, thereby realizing various functions, as follows:

[0303] Obtain object sample data in a target object recognition scenario, perform data enhancement processing on the object sample data, and obtain enhanced object sample data;

[0304] Perform feature extraction on the enhanced object sample data through a plurality of different object recognition models, and obtain homologous object features extracted by each object recognition model;

[0305] Based on the homologous object features extracted by each object recognition model, obtain a robustness measurement score corresponding to each object recognition model;

[0306] Based on the reference object feature information extracted by each object recognition model for the object sample data, obtain an accuracy measurement score corresponding to each object recognition model;

[0307] According to the accuracy measurement score and the robustness measurement score of each object recognition model, filter out a target object recognition model that adapts to the target object recognition scenario from the plurality of different object recognition models, so as to perform object recognition through the target object recognition model.

[0308] The specific implementation of each operation can refer to the foregoing embodiments, which will not be described here.

[0309] According to an aspect of the present application, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method provided in various optional implementation manners in the above embodiments.

[0310] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by a computer program, or by a computer program controlling related hardware, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0311] To this end, the embodiment of the present application also provides a storage medium, wherein the computer program is stored, the computer program can be loaded by the processor to execute the steps in any object recognition method provided by the embodiment of the present application. For example, the computer program can execute the following steps:

[0312] Obtain object sample data in a target object recognition scene, perform data enhancement processing on the object sample data, and obtain enhanced object sample data;

[0313] Extract features from the enhanced object sample data through a plurality of different object recognition models, and obtain homologous object features extracted by each object recognition model;

[0314] Based on the homologous object features extracted by each object recognition model, obtain a robustness measurement score corresponding to each object recognition model;

[0315] Based on the reference object feature information extracted by each object recognition model for the object sample data, obtain an accuracy measurement score corresponding to each object recognition model;

[0316] According to the accuracy measurement score and the robustness measurement score of each object recognition model, filter a target object recognition model from the plurality of different object recognition models, and perform object recognition through the target object recognition model.

[0317] The specific implementation of each operation can be referred to the foregoing embodiments, which will not be repeated here.

[0318] Since the computer program stored in the storage medium can execute the steps in any object recognition method provided by the embodiment of the present application, the beneficial effects of any object recognition method provided by the embodiment of the present application can be achieved, which will be described in detail in the foregoing embodiments, and will not be repeated here.

[0319] The above describes in detail the object recognition method, device, computer device and storage medium provided by the embodiment of the present application. The specific examples are applied to explain the principle and implementation mode of the present application. The above embodiment is only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed; in conclusion, the content of the specification should not be understood as the limitation of the present application.

Claims

1. An object recognition method, characterized in that, include: Obtain object sample data in a target object recognition scenario, and perform data augmentation processing on the object sample data to obtain augmented object sample data; The enhanced object sample data is subjected to feature extraction by multiple different object recognition models to obtain the same object features extracted by each object recognition model; wherein, the same object features refer to the features obtained by feature extraction of the same object sample data, and the same object sample data refers to the data from the same object identifier; Based on the features of the same object extracted by each object recognition model, obtain the robustness score corresponding to each object recognition model; Based on the benchmark object feature information extracted by each object recognition model from the object sample data, an accuracy measurement score is obtained for each object recognition model; wherein, the benchmark object feature information includes the object features of the objects in the object sample data; Based on the accuracy and robustness scores of each object recognition model, a target object recognition model suitable for the target object recognition scenario is selected from the multiple different object recognition models, so as to perform object recognition through the target object recognition model.

2. The method according to claim 1, characterized in that, The process of obtaining an accuracy score for each object recognition model based on the baseline object feature information extracted from the object sample data by each object recognition model includes: The baseline object feature information extracted by each object recognition model is used to identify the accuracy reference threshold corresponding to each object recognition model; Based on the label information corresponding to the object sample data, a recall threshold list corresponding to each object recognition model is identified from the baseline object feature information of each object recognition model. Based on the accuracy reference threshold corresponding to each object recognition model, the recall threshold list corresponding to each object recognition model is processed to obtain the accuracy measurement score corresponding to each object recognition model.

3. The method according to claim 2, characterized in that, The process of identifying the baseline object feature information extracted by each object recognition model to obtain the accuracy reference threshold corresponding to each object recognition model includes: The object sample data are combined to obtain an object sample data group; Based on the baseline object feature information, calculate the object spatial feature similarity between object sample data in the object sample data group; Based on the label information of the object sample data in the object sample data group, a set of heterogeneous image groups is identified in the object sample data group; Based on the similarity of object spatial features corresponding to the object sample data groups in the heterogeneous image group set, an accuracy reference threshold corresponding to the object recognition model is generated.

4. The method according to claim 2, characterized in that, The step involves processing the recall threshold list corresponding to each object recognition model based on the accuracy reference threshold corresponding to each object recognition model to obtain the accuracy measurement score corresponding to each object recognition model, including: The object spatial feature similarity in the recall threshold list is compared with the accuracy reference threshold to obtain the comparison result; Based on the comparison results, the similarity of misjudged features is determined in the recall threshold list; The accuracy score of the object recognition model is obtained by comparing the number of misjudged feature similarities with the number of object spatial feature similarities in the recall threshold list.

5. The method according to claim 1, characterized in that, The process of obtaining a robustness score for each object recognition model based on the features of shared objects extracted by each object recognition model includes: The features of multiple homologous objects identified by the object recognition model are combined to obtain a homologous object feature group. Calculate the similarity distance between features of similar objects in a feature group; The similarity distance of the feature groups of the same objects is denoised to obtain the feature similarity of the same objects in the object recognition model. Based on the similarity of features of the same source objects corresponding to each object recognition model, a robustness score is determined for each object recognition model.

6. The method according to claim 1, characterized in that, The process of performing data augmentation on the object sample data to obtain augmented object sample data includes: Acquire a variety of different data augmentation methods; Determine the augmentation parameters for each data augmentation method; Based on the enhancement parameters corresponding to each data augmentation method, multiple different data augmentation methods are used to perform data augmentation processing on the object sample data to obtain the augmented object sample data.

7. The method according to claim 1, characterized in that, Before extracting features from the enhanced object sample data using multiple different object recognition models to obtain the homologous object features extracted by each object recognition model, the method further includes: Obtain the object recognition model to be trained and the object training sample data; The spatial features of the object training sample data are extracted using the object recognition model to be trained, thereby obtaining the spatial structure information corresponding to the object training sample data. The spatial structure information corresponding to the object training sample data is processed by a fully connected mapping using the object recognition model to be trained, and the recognition result corresponding to the object training sample data is obtained. Based on the preset label information of the object training sample data and the recognition result corresponding to the object training sample data, calculate the model loss information; Based on the model loss information, the model parameters of the object recognition model to be trained are adjusted to obtain the object recognition model.

8. An object recognition device, characterized in that, include: The data augmentation unit is used to acquire object sample data in the target object recognition scenario, perform data augmentation processing on the object sample data, and obtain augmented object sample data. The feature extraction unit is used to extract features from the enhanced object sample data using multiple different object recognition models to obtain the same object features extracted by each object recognition model; wherein, the same object features refer to the features obtained by feature extraction from the same object sample data, and the same object sample data refers to the data from the same object identifier; The first score acquisition unit is used to acquire the robustness measurement score corresponding to each object recognition model based on the features of the same object extracted by each object recognition model. The second score acquisition unit is used to acquire the accuracy measurement score corresponding to each object recognition model based on the benchmark object feature information extracted by each object recognition model for the object sample data; wherein, the benchmark object feature information includes the object features of the objects in the object sample data; The filtering unit is used to filter out the target object recognition model that is suitable for the target object recognition scenario from the multiple different object recognition models based on the accuracy measurement score and robustness measurement score of each object recognition model, so as to perform object recognition through the target object recognition model.

9. The apparatus according to claim 8, characterized in that, The second score acquisition unit includes: The first identification subunit is used to identify the baseline object feature information extracted by each object identification model and obtain the accuracy reference threshold corresponding to each object identification model. The second identification subunit is used to identify the recall threshold list corresponding to each object identification model in the baseline object feature information of each object identification model based on the label information corresponding to the object sample data. The first discrimination subunit is used to discriminate the recall threshold list corresponding to each object recognition model based on the accuracy reference threshold corresponding to each object recognition model, and obtain the accuracy measurement score corresponding to each object recognition model.

10. The apparatus according to claim 9, characterized in that, The first identification subunit includes: The combination module is used to combine the object sample data to obtain an object sample data group; The calculation module is used to calculate the object spatial feature similarity between object sample data in the object sample data group based on the benchmark object feature information; The identification module is used to identify a set of heterogeneous image groups in the object sample data group based on the label information of the object sample data in the object sample data group. The generation module is used to generate an accuracy reference threshold for the object recognition model based on the object spatial feature similarity corresponding to the object sample data group in the heterogeneous image group set.

11. The apparatus according to claim 9, characterized in that, The discrimination subunit includes: The comparison module is used to compare the spatial feature similarity of objects in the recall threshold list with the accuracy reference threshold to obtain a comparison result; The determination module is used to determine the similarity of misjudged features in the recall threshold list based on the comparison results; The comparison calculation module is used to compare the number of misjudged feature similarities with the number of object spatial feature similarities in the recall threshold list to obtain the accuracy measurement score corresponding to the object recognition model.

12. The apparatus according to claim 8, characterized in that, The first score acquisition unit includes: The combination subunit is used to combine multiple features of the same object identified by the object recognition model to obtain a feature group of the same object; The calculation subunit is used to calculate the similarity distance between features of similar objects in a feature group of similar objects; The denoising subunit is used to denoise the similarity distance of the feature groups of the same objects to obtain the feature similarity of the same objects in the object recognition model. The second discrimination subunit is used to determine the robustness score of each object recognition model based on the similarity of features of the same source object corresponding to each object recognition model.

13. The apparatus according to claim 8, characterized in that, The data enhancement unit includes: Get sub-units to acquire various data augmentation methods; The parameter determination subunit is used to determine the enhancement parameters corresponding to each data augmentation method. The data augmentation subunit is used to perform data augmentation processing on the object sample data using multiple different data augmentation methods based on the augmentation parameters corresponding to each data augmentation method, so as to obtain the augmented object sample data.

14. The apparatus according to claim 8, characterized in that, The object recognition device further includes: The acquisition unit is used to acquire the object recognition model to be trained and the object training sample data. The feature extraction unit is used to extract spatial features from the object training sample data using the object recognition model to be trained, so as to obtain the spatial structure information corresponding to the object training sample data. The fully connected mapping unit is used to perform fully connected mapping processing on the spatial structure information corresponding to the object training sample data using the object recognition model to be trained, so as to obtain the recognition result corresponding to the object training sample data. The computing unit is used to calculate the model loss information based on the preset label information of the object training sample data and the recognition results corresponding to the object training sample data. The adjustment unit is used to adjust the model parameters of the object recognition model to be trained based on the model loss information, so as to obtain the object recognition model.

15. A computer device, characterized in that, It includes a memory and a processor; the memory stores an application program, and the processor runs the application program within the memory to perform the operations in the object recognition method according to any one of claims 1 to 7.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the object recognition method according to any one of claims 1 to 7.

17. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium; the processor of the computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the object recognition method according to any one of claims 1 to 7.

Images