Object feature extraction model generation method and object recognition method

Abstract

The present disclosure relates to a method for generating an object feature extraction model and an object recognition method. The method comprises: obtaining a plurality of to-be-integrated models from a plurality of object feature extraction models; performing feature extraction performance testing on the plurality of to-be-integrated models to obtain a first test result corresponding to each of the plurality of to-be-integrated models; selecting a current integrated model corresponding to a target test result from the plurality of to-be-integrated models; performing model integration processing on the current integrated model and a preset integrated model to obtain a target integrated model; performing extraction performance comparison processing on the target integrated model and the preset integrated model to obtain a target comparison result; and generating a target object feature extraction model based on the preset integrated model in a case where the target comparison result indicates that the object feature extraction performance of the preset integrated model is superior to the object feature extraction performance of the target integrated model. The embodiments of the present disclosure can improve the object feature extraction performance of the target object feature extraction model and improve the model deployment efficiency.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a method for generating an object feature extraction model and an object recognition method. Background Technology

[0002] Currently, with the widespread application of mobile facial recognition systems in high-security scenarios such as security, payment, and access control, the accuracy requirements for facial feature extraction models are continuously increasing. Model ensemble learning is a stable and effective method to improve model recognition accuracy. Model ensemble learning combines multiple models with weaker object feature extraction performance. These weaker classifiers undergo complementarity screening and testing before being combined to obtain a better-performing model. However, model ensemble learning integrates features obtained from multiple weaker classifications, requiring forward inference operations on multiple models simultaneously during deployment. This forward inference increases the total deployment time and consumes computational resources, hindering the deployment and application of models in real-time scenarios. Mobile applications have strict requirements regarding model latency; using multiple models for object feature extraction results in deployment times that cannot meet these requirements. Summary of the Invention

[0003] In view of the above-mentioned technical problems, this disclosure proposes a method for generating an object feature extraction model and an object recognition method.

[0004] According to one aspect of the embodiments of this disclosure, a method for generating an object feature extraction model is provided, comprising:

[0005] Multiple models to be integrated are obtained from multiple object feature extraction models, wherein the multiple models to be integrated are object feature extraction models that have not participated in the integration process among the multiple object feature extraction models;

[0006] The feature extraction performance of the plurality of models to be integrated is tested to obtain a first test result corresponding to the plurality of models to be integrated; the first test result characterizes the object feature extraction performance of each of the plurality of models to be integrated.

[0007] The current integrated model corresponding to the target test result is selected from the plurality of models to be integrated; the target test result is the test result with the best performance in extracting the indicator object features among the first test results;

[0008] The current ensemble model and the preset ensemble model are subjected to model ensemble processing to obtain the target ensemble model. The preset ensemble model is obtained by ensemble of at least one of the multiple object feature extraction models whose object feature extraction performance is better than that of the current ensemble model.

[0009] The target ensemble model and the preset ensemble model are subjected to performance extraction comparison processing to obtain the target comparison result;

[0010] If the target comparison result indicates that the object feature extraction performance of the preset ensemble model is better than that of the target ensemble model, a target object feature extraction model is generated based on the preset ensemble model.

[0011] According to another aspect of the embodiments of this disclosure, an object recognition method is provided, including:

[0012] Acquire an image of the object to be identified and at least one reference object image;

[0013] The image of the object to be identified and the at least one reference object image are input into the target object feature extraction model for feature extraction to obtain the third object feature information corresponding to the image of the object to be identified and the fourth object feature information corresponding to the at least one reference object image;

[0014] Based on the third object feature information and the fourth object feature information, object recognition is performed on the target object in the image to be identified, and the object recognition result is obtained.

[0015] The target object feature extraction model is obtained based on the above-mentioned object feature extraction model generation method.

[0016] According to another aspect of the embodiments of this disclosure, an apparatus for generating an object feature extraction model is provided, comprising:

[0017] The module for obtaining models to be integrated is used to obtain multiple models to be integrated from multiple object feature extraction models, wherein the models to be integrated are object feature extraction models that have not participated in the integration process among the multiple object feature extraction models.

[0018] The feature extraction performance testing module is used to perform feature extraction performance testing on the plurality of models to be integrated, and obtain a first test result corresponding to the plurality of models to be integrated; the first test result characterizes the object feature extraction performance of each of the plurality of models to be integrated.

[0019] The current ensemble model determination module is used to select the current ensemble model corresponding to the target test result from the plurality of models to be integrated; the target test result is the test result with the best performance in the first test result indicating the object feature extraction performance.

[0020] The model integration module is used to perform model integration processing on the current integrated model and the preset integrated model to obtain a target integrated model. The preset integrated model is obtained by integrating at least one of the multiple object feature extraction models whose object feature extraction performance is better than the current integrated model.

[0021] An extraction performance comparison module is used to perform extraction performance comparison processing on the target ensemble model and the preset ensemble model to obtain the target comparison result;

[0022] The target object feature extraction model determination module is used to generate a target object feature extraction model based on the preset integrated model when the target comparison result indicates that the object feature extraction performance of the preset integrated model is better than that of the target integrated model.

[0023] According to another aspect of the embodiments of this disclosure, an object recognition device is provided, comprising:

[0024] The image acquisition module is used to acquire an image of the object to be identified and at least one reference object image;

[0025] The feature extraction module is used to input the image of the object to be identified and the at least one reference object image into the target object feature extraction model for feature extraction, so as to obtain the third object feature information corresponding to the image of the object to be identified and the fourth object feature information corresponding to the at least one reference object image; the target object feature extraction model is obtained based on the above-described object feature extraction model generation method;

[0026] The object recognition module is used to perform object recognition on the target object in the image to be recognized based on the third object feature information and the fourth object feature information, and obtain the object recognition result.

[0027] According to another aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the above-described object feature extraction model generation method or the above-described object recognition method.

[0028] According to another aspect of the present disclosure, a computer-readable storage medium is provided, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the above-described object feature extraction model generation method or the above-described object recognition method.

[0029] According to another aspect of the present disclosure, a computer program product containing instructions is provided, which, when run on a computer, causes the computer to execute the above-described object feature extraction model generation method or the above-described object recognition method.

[0030] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0031] By obtaining multiple unintegrated models from multiple object feature extraction models, and performing feature extraction performance tests on these models, first test results representing the object feature extraction performance of each model are obtained. The current integrated model corresponding to the target test result is then selected from the multiple models. This current integrated model is then integrated with a preset integrated model to obtain the target integrated model. The preset integrated model is obtained by integrating at least one model whose object feature extraction performance is superior to the current integrated model. By integrating the current integrated model with the preset integrated model, the object feature extraction performance of the target integrated model can be improved. The performance of the target ensemble model and the preset ensemble model is compared to obtain the target comparison results. If the target comparison results indicate that the object feature extraction performance of the preset ensemble model is better than that of the target ensemble model, the target object feature extraction model is generated based on the preset ensemble model. This can improve the object feature extraction performance of the target object feature extraction model and increase the distinctiveness of the features extracted by the target object feature extraction model. Moreover, the target object feature extraction model is essentially a single object feature extraction model. Compared with the deployment process of multiple models, the forward inference time of the target object feature extraction model can be greatly shortened, thereby improving the model deployment efficiency and making it suitable for a variety of complex application scenarios.

[0032] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0033] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0034] Figure 1 This is a schematic diagram illustrating an application system according to an exemplary embodiment;

[0035] Figure 2 This is a flowchart illustrating a method for generating an object feature extraction model according to an exemplary embodiment;

[0036] Figure 3 This is a schematic diagram illustrating the update process of model integration in a method for generating an object feature extraction model according to an exemplary embodiment;

[0037] Figure 4This is a block diagram illustrating an apparatus for generating an object feature extraction model according to an exemplary embodiment;

[0038] Figure 5 This is a block diagram illustrating an electronic device for generating a target object feature extraction model or identifying a target object, according to an exemplary embodiment.

[0039] Figure 6 This is a block diagram illustrating another electronic device for generating a target object feature extraction model or identifying a target object, according to an exemplary embodiment. Detailed Implementation

[0040] Various exemplary embodiments, features, and aspects of this application will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0041] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0042] Furthermore, to better illustrate this application, numerous specific details are provided in the following detailed embodiments. Those skilled in the art should understand that this application can be implemented without certain specific details. In some instances, methods, means, components, and circuits well-known to those skilled in the art have not been described in detail in order to highlight the main points of this application.

[0043] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or computers-controlled machines to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. AI software technology mainly includes computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0044] In recent years, with the research and progress of artificial intelligence technology, it has been widely applied in many fields. The solutions provided in this application involve technologies such as machine learning / deep learning, which are specifically illustrated through the following embodiments:

[0045] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating an application system according to an exemplary embodiment. The application system can be used in the object feature extraction model generation method or object recognition method of this application. Figure 1 As shown, the application system may include at least server 01 and terminal 02.

[0046] In this embodiment, server 01 can be used to generate a feature extraction model for the target object. Specifically, the server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0047] In this embodiment, terminal 02 can be a terminal used for object recognition. The aforementioned user terminal may include physical devices such as smartphones, desktop computers, tablets, laptops, smart speakers, in-vehicle terminals, digital assistants, augmented reality (AR) / virtual reality (VR) devices, and smart wearable devices, and may also include software running on the physical device, such as applications. The operating system running on the aforementioned user terminal in this embodiment may include, but is not limited to, Android, iOS, Linux, and Windows systems.

[0048] In addition, it should be noted that, Figure 1 The example shown is merely one application environment provided by this disclosure. In practical applications, other application environments may also be included. For example, the identification of the target object may also be implemented on the server.

[0049] In the embodiments described in this specification, the terminal 02 and the server 01 can be directly or indirectly connected through wired or wireless communication, and this application does not limit this connection.

[0050] It should be noted that the following diagram shows one possible sequence of steps, and it is not strictly necessary to follow this order. Some steps can be executed in parallel without interdependence.

[0051] Specifically, Figure 2 This is a flowchart illustrating a method for generating an object feature extraction model according to an exemplary embodiment. For example... Figure 2 As shown, the method for generating this object feature extraction model can be used in electronic devices such as terminals or servers, and may include the following steps:

[0052] In step S201, multiple models to be integrated are obtained from multiple object feature extraction models.

[0053] In one specific embodiment, each object feature extraction model can be used to extract features of objects in the image to be identified. It is understood that the object feature extraction model can extract spatial features from the image to be identified, obtaining feature information that characterizes the spatial structure of objects in the image. Specifically, the objects in the image to be identified may include faces or hands, for example, a human face. Optionally, the model structure of the object feature extraction model can be a convolutional neural network model structure, or other model structures, which are not limited in this disclosure; wherein, the convolutional neural network model structure may specifically include operations such as convolution calculation, nonlinear activation function calculation, and pooling calculation.

[0054] In one specific embodiment, the multiple models to be integrated can be object feature extraction models that have not participated in the integration process among multiple object feature extraction models.

[0055] In one specific embodiment, the above method may further include:

[0056] Obtain the sample object image set, the preset category information corresponding to the sample object image set, and the target category feature information corresponding to the sample object image set;

[0057] The first current training image set and the first preset category information corresponding to the first current training image set are determined from the sample object image set and the preset category information, respectively.

[0058] The first current training image set is input into the first preset feature extraction model to extract features, thereby obtaining the first object feature information corresponding to the first current training image set;

[0059] Based on the first object feature information and the target category feature information, determine the first predicted category information corresponding to the first current training image set;

[0060] Based on the first predicted category information and the first preset category information, the first loss information is determined;

[0061] Based on the first loss information, the first preset feature extraction model is updated. Based on the updated first preset feature extraction model, the training iteration steps of determining the first current training image set and the first preset category information corresponding to the first current training image set from the sample object image set and the preset category information are repeated until the first preset convergence condition is reached.

[0062] The trained feature extraction model is determined based on the first preset feature extraction model obtained when the first preset convergence condition is met.

[0063] In one specific embodiment, the sample object image set can be used to train a first preset feature extraction model, a second preset feature extraction model, and preset category feature information. The sample object image set may include multiple sample object images. A sample object can refer to the object to be identified in the sample object image; a sample object may include a face or hand, for example, a human face. The preset category information corresponding to the sample object image set may include object category information corresponding to each of the multiple sample object images. The object category information corresponding to each sample object image can characterize the object category corresponding to the sample object in that sample object image.

[0064] In one specific embodiment, the target category feature information may include the category center information corresponding to each of multiple sample objects in the sample object image set. Specifically, the target category feature information may be represented as a matrix or multiple vectors, etc.; in this embodiment, the target category feature information may be a d*k matrix, where d is the feature dimension, k is the number of sample objects in the sample object image set, and each row in the matrix can be the category center information corresponding to each sample object in the sample object image set. The category center information of each sample object can be a class center used to characterize the sample object; specifically, the class center may be represented as a vector, etc.

[0065] In one specific embodiment, the above-mentioned target category feature information is obtained using the following method:

[0066] Obtain preset category feature information;

[0067] The second current training image set and the second preset category information corresponding to the second current training image set are determined from the sample object image set and the preset category information, respectively;

[0068] The second current training image set is input into the second preset feature extraction model to extract features, thereby obtaining the feature information of the second object corresponding to the second current training image set;

[0069] Based on the second object feature information and the preset category feature information, the second predicted category information corresponding to the second current training image set is determined;

[0070] Based on the second predicted category information and the second preset category information, the second loss information is determined;

[0071] Based on the second loss information, the preset category feature information and the second preset feature extraction model are updated. Based on the updated preset category feature information and the updated second preset feature extraction model, the iterative steps of determining the second current training image set and the second preset category information corresponding to the second current training image set from the sample object image set and the preset category information are repeated until the second preset convergence condition is reached.

[0072] The preset category feature information obtained when the second preset convergence condition is met is used as the target category feature information.

[0073] In a specific embodiment, the preset category feature information can refer to the object category feature information to be trained. The preset category feature information can characterize the object category features corresponding to each of multiple sample objects in the sample object image set. The preset category feature information can be pre-set; optionally, it can be obtained by randomly initializing a corresponding number of values ​​based on the number of sample objects in the sample object image set. Specifically, the preset category feature information can be a d*k matrix, where d is the feature dimension and k is the number of sample objects in the sample object image set; the category center information corresponding to each sample object in the sample object image set can be a vector of dimension d.

[0074] In one specific embodiment, the second current training image set can refer to the set of sample object images currently used to train the second preset feature extraction model during multiple training processes. The second preset category information can include object category information corresponding to each of the multiple sample object images in the second current training image set. Specifically, the object category information corresponding to any sample object image can characterize the sample object to which that sample object image belongs. The object category information corresponding to any sample object image can be the identification information of the sample object to which that sample object image belongs; for example, the identification information of the sample object can be represented in numerical form.

[0075] In a specific embodiment, a preset number of sample object images can be randomly selected from the sample object image set as the second current training image set, and the second preset category information corresponding to the second current training image set can be obtained based on the object category information corresponding to each sample object image in the preset category information.

[0076] In one specific embodiment, the second preset feature extraction model can refer to a model to be trained and used to extract object features from an image. The second object feature information can include object feature information corresponding to multiple sample object images in the second current training image set. The object feature information corresponding to each sample object image in the second current training image set can characterize the features of the sample object in that sample object image.

[0077] In one specific embodiment, feature extraction can be performed on each sample object image in the second current training image set using a second preset feature extraction model to obtain the second object feature information.

[0078] In one specific embodiment, the second predicted category information can characterize the probability that any sample object image in the second current training image set belongs to any object category. Specifically, the second predicted category information can include a probability vector of any sample object image in the second current training image set belonging to any object category.

[0079] In one specific embodiment, the second predicted category information can be obtained by performing matrix multiplication on the second object feature information and the preset category feature information. For example, the second predicted category information C can be calculated using the following formula:

[0080] C = A × B

[0081] Where A represents the second object feature information, and B represents the preset category feature information.

[0082] In one specific embodiment, the second loss information can characterize the deviation between the second predicted category information and the second preset category information. Specifically, the second loss information can be calculated using a loss function based on the second predicted category information and the second preset category information. Optionally, the loss function can be a classification function, or other types of loss functions can be used; for example, the classification function can include a normalized exponential function or a normalized exponential function with boundary constraints, etc.

[0083] In one specific embodiment, the model parameters and preset category feature information in the second preset feature extraction model can be updated based on the second loss information. Specifically, an update gradient can be determined based on the second loss information, and the model parameters and preset category feature information in the second preset feature extraction model can be updated based on this update gradient. In this embodiment, the parameters and preset category feature information in the network structure such as convolutional layers, normalization layers, and nonlinear layers in the second preset feature extraction model can be updated. The second preset feature extraction model can be trained and optimized based on gradient descent, for example, stochastic gradient descent or stochastic gradient descent with moving averages, etc., which is not limited in this disclosure.

[0084] In one specific embodiment, the second preset convergence condition may be that the second loss information is less than a preset loss threshold, or it may be that the number of training iterations meets a preset number of iterations, which is not limited here.

[0085] In the above embodiments, by training the preset category feature information and the second preset feature extraction model with the sample object image set, the target category feature information corresponding to the sample object image set can be obtained. Then, in the process of training the first preset feature extraction model with the sample object image set to obtain multiple object feature extraction models, the same target category feature information can be used for training. On this basis, since the feature dimensions used in the model are the same and the features corresponding to each dimension are the same, the feature space of the model can be constrained to a unified spherical distribution. There is no need to train a large number of feature extraction sub-models for combination testing, which can improve training efficiency and greatly shorten the training cycle. Furthermore, by constraining the feature space of the model to a unified spherical distribution, the efficiency of obtaining effective complementary models can be improved.

[0086] In a specific embodiment, the first current training image set may refer to the set of sample object images currently training the first preset feature extraction model during multiple iterations of training.

[0087] In one specific embodiment, the first preset category information may include object category information corresponding to each of the multiple sample object images in the first current training image set. Specifically, the object category information corresponding to any sample object image may characterize the sample object to which that sample object image belongs. The object category information corresponding to any sample object image may be the identification information of the sample object to which that sample object image belongs; for example, the identification information of the sample object may be represented in the form of a numerical value.

[0088] In a specific embodiment, a preset number of sample object images can be randomly selected from the sample object image set as the first current training image set, and the first preset category information corresponding to the first current training image set can be obtained based on the object category information corresponding to each sample object image in the preset category information.

[0089] In one specific embodiment, the first preset feature extraction model can refer to a model to be trained and used to extract object features from an image. The first object feature information can include object feature information corresponding to multiple sample object images in the first current training image set. The object feature information corresponding to each sample object image in the first current training image set can characterize the features of the sample object in that sample object image.

[0090] In a specific embodiment, feature extraction can be performed on each sample object image in the first current training image set using a first preset feature extraction model to obtain the first object feature information.

[0091] In one specific embodiment, the first predicted category information can characterize the probability that any sample object image in the first current training image set belongs to any object category. Specifically, the first predicted category information may include a probability vector of any sample object image in the first current training image set belonging to any object category.

[0092] In one specific embodiment, the first predicted category information can be obtained by performing matrix multiplication on the first object feature information and the target category feature information. For example, the first predicted category information F can be calculated using the following formula:

[0093] F = D × E

[0094] Where D represents the first object feature information, and E represents the target category feature information.

[0095] In a specific embodiment, the first loss information can characterize the deviation between the first predicted category information and the first preset category information. Specifically, the first loss information can be calculated using a loss function based on the first predicted category information and the first preset category information. Optionally, the loss function can be a classification function, or other types of loss functions can be used; for example, the classification function can include a normalized exponential function or a normalized exponential function with boundary constraints, etc.

[0096] In a specific embodiment, when the loss function is a loss function with boundary constraints, the loss function can be:

[0097]

[0098] Where N is the number of sample object images in the first current training image set, θ yi θ represents the directional information between the object feature information corresponding to the i-th sample object image in the first object feature information and the category feature information corresponding to the object feature information in the target category feature information. j The directional information between any two category features that do not correspond to the above object features in the target category feature information, k is the number of sample objects in the sample object image set, and y is the direction information between any two category features. i Here, s, m1, and m2 represent the preset category information corresponding to the i-th sample object image in the first current training image set, and m1, m2 are preset parameters. It is understood that adjusting s, m1, and m2 to ensure the diversity of spatial constraints of the images can promote the model's learning of complementary knowledge. In the above embodiment, by controlling the settings of s, m1, and m2, different face recognition models to be integrated can be obtained, thereby improving the success rate of acquiring complementary object feature acquisition models.

[0099] In one specific embodiment, the model parameters in the first preset feature extraction model can be updated based on the first loss information. Specifically, an update gradient can be determined based on the first loss information, and the model parameters in the first preset feature extraction model can be updated based on this update gradient. In this embodiment, the parameters in the network structure such as convolutional layers, normalization layers, and nonlinear layers in the first preset feature extraction model can be updated. The first preset feature extraction model can be trained and optimized based on gradient descent; for example, it can be stochastic gradient descent or stochastic gradient descent with moving averages, etc., which is not limited in this disclosure.

[0100] In a specific embodiment, the first preset convergence condition may be that the first loss information is less than a preset loss threshold, or it may be that the number of training iterations meets a preset number of iterations, which is not limited here.

[0101] In one specific embodiment, the trained feature extraction model can be any one of multiple object feature extraction models. It is understood that a trained feature extraction model can be obtained by training the first preset feature extraction model each time; correspondingly, multiple different trained feature extraction models can be obtained through successive training. Optionally, the second preset feature extraction model that reaches the second preset convergence condition can be used as the trained feature extraction model.

[0102] In the above embodiments, by using the same target category feature information for training the first preset feature extraction model to obtain multiple object feature extraction models, the feature space of the model can be constrained to a unified spherical distribution. This eliminates the need to train a large number of feature extraction sub-models in different feature spaces for combined testing, thereby improving training efficiency and greatly shortening the training cycle. Furthermore, by constraining the feature space of the model to a unified spherical distribution, the efficiency of obtaining effective complementary models can be improved.

[0103] In one specific embodiment, step S201 above may include:

[0104] Identify at least one integrated object feature extraction model from a plurality of object feature extraction models;

[0105] At least one integrated object feature extraction model is removed from the multiple object feature extraction models to obtain multiple models to be integrated.

[0106] In a specific embodiment, at least one integrated object feature extraction model can refer to a model that has participated in model integration processing among multiple object feature extraction models.

[0107] In step S203, feature extraction performance tests are performed on multiple models to be integrated to obtain the first test results corresponding to the multiple models to be integrated.

[0108] In one specific embodiment, the first test result can characterize the object feature extraction performance of each of the multiple models to be integrated.

[0109] In one specific embodiment, feature extraction performance tests can be performed on multiple models to be integrated based on a test image set. The test image set may include multiple sample object images.

[0110] Specifically, taking any one of the multiple models to be integrated as an example, by inputting the sample object image set into the model to be integrated, object feature information corresponding to multiple sample object images in the sample object image set can be obtained; based on the object feature information corresponding to multiple sample object images in the sample object image set, feature comparison is performed on the object feature information corresponding to each two different sample object images to obtain object association information corresponding to multiple comparison combinations, wherein any comparison combination can include any two sample object images from the multiple sample object images; sorting the object association information corresponding to multiple identical sample object combinations in descending order of association degree, an identical sample object comparison sequence can be obtained, wherein an identical sample object combination can refer to the sample object images in any comparison combination of multiple comparison combinations belonging to the same sample object. For combinations, the object association information corresponding to multiple combinations of dissimilar sample objects is sorted in descending order of association degree to obtain a dissimilar sample object comparison sequence. A dissimilar sample object combination can refer to a comparison combination where the sample object images within any of the multiple comparison combinations belong to different sample objects. Based on a preset error rate and the aforementioned dissimilar sample object comparison sequence, a reference association information threshold can be determined. Based on the same sample object comparison sequence and the reference association information threshold, the number of associated combinations can be obtained. The number of associated combinations can refer to the number of same sample object combinations in the same sample object comparison sequence that are greater than or equal to the aforementioned reference association information threshold. Based on the number of associated combinations and the number of same sample object combinations, the pass rate of the model to be integrated can be obtained, and correspondingly, this pass rate can be used as the first test result of the model to be integrated. It is understood that, with the same preset error rate, the higher the pass rate of the model to be integrated, the better the object feature extraction performance of the model to be integrated. Optionally, the object association information can be feature similarity.

[0111] In step S205, the current integration model corresponding to the target test result is selected from multiple models to be integrated.

[0112] In one specific embodiment, the target test result can be the test result that indicates the optimal performance of object feature extraction in the first test result.

[0113] In a specific embodiment, the performance sequence of the models to be integrated can be obtained by sorting the test results corresponding to the multiple models to be integrated in the first test result according to the performance of object feature extraction from best to worst; the first test result in the performance sequence of the models to be integrated is taken as the target test result; and the model to be integrated corresponding to the target test result is taken as the current integrated model.

[0114] In step S207, the current integrated model and the preset integrated model are subjected to model integration processing to obtain the target integrated model.

[0115] In one specific embodiment, the preset ensemble model can be obtained by integrating at least one model from multiple object feature extraction models whose object feature extraction performance is superior to the current ensemble model. It is understood that the preset ensemble model can be obtained through at least one model ensemble process.

[0116] In a specific embodiment, the model parameters of the target ensemble model can be obtained by averaging the corresponding model parameters of the current ensemble model and the preset ensemble model. It is understood that multiple object feature extraction models are obtained by training a first preset feature extraction model, and these multiple object feature extraction models have the same model structure. The target ensemble model can be obtained by integrating the model parameters. Specifically, taking the model parameters of the convolutional layers in a convolutional neural network model structure as an example, assuming that the convolutional layers of the current ensemble model may include a first convolutional kernel W1 and a second convolutional kernel W2, and correspondingly, the convolutional layers of the preset ensemble model may include a third convolutional kernel W3 corresponding to the first convolutional kernel W1 and a fourth convolutional kernel W4 corresponding to the second convolutional kernel W2, the fifth convolutional kernel W5 and the sixth convolutional kernel W6 in the convolutional layers of the target ensemble model can be calculated using the following formula:

[0117]

[0118]

[0119] Wherein, W1 is the first convolutional kernel in the convolutional layer of the current ensemble model, W2 is the second convolutional kernel in the convolutional layer of the current ensemble model, W3 is the third convolutional kernel in the convolutional layer of the preset ensemble model, and W4 is the fourth convolutional kernel in the convolutional layer of the preset ensemble model.

[0120] In step S209, the target ensemble model and the preset ensemble model are subjected to performance comparison processing to obtain the target comparison result.

[0121] In one specific embodiment, the target alignment result may include a first alignment result and a second alignment result. The first alignment result may indicate that the object feature extraction performance of the preset ensemble model is better than that of the target ensemble model; the second alignment result may indicate that the object feature extraction performance of the target ensemble model is better than that of the preset ensemble model.

[0122] In one specific embodiment, step S209 above may include:

[0123] The feature extraction performance of the target ensemble model and the preset ensemble model is tested to obtain the second test result corresponding to the target ensemble model and the third test result corresponding to the preset ensemble model.

[0124] The results of the second and third tests are compared to obtain the target comparison results.

[0125] In one specific embodiment, the second test result can characterize the object feature extraction performance of the target ensemble model. The third test result can characterize the object feature extraction performance of the preset ensemble model.

[0126] In a specific embodiment, the above feature extraction performance testing process can refer to step S203 above, and will not be repeated here.

[0127] In one specific embodiment, the pass rate in the second test result can be compared with the pass rate in the third test result to obtain the target comparison result.

[0128] In step S211, if the target comparison result indicates that the object feature extraction performance of the preset ensemble model is better than that of the target ensemble model, a target object feature extraction model is generated based on the preset ensemble model.

[0129] In one specific embodiment, the target object feature extraction model can be used to extract object features from an image. The target object feature extraction model can be applied to various object recognition tasks; for example, the object can be a person, and the object recognition task can include, but is not limited to, face recognition or human body recognition tasks.

[0130] In one specific embodiment, if the target comparison result indicates that the object feature extraction performance of the preset ensemble model is better than that of the target ensemble model, the preset ensemble model can be used as the target object feature extraction model.

[0131] In the above embodiments, multiple unintegrated models are obtained from multiple object feature extraction models. Feature extraction performance tests are performed on these models to obtain first test results representing the object feature extraction performance of each model. A current integrated model corresponding to the target test result is selected from the multiple models. This current integrated model is then integrated with a preset integrated model to obtain the target integrated model. The preset integrated model is obtained by integrating at least one model whose object feature extraction performance is superior to the current integrated model. By integrating the current integrated model with the preset integrated model, the object feature extraction performance of the target integrated model can be improved. The feature extraction performance is then compared between the target ensemble model and the preset ensemble model to obtain the target comparison result. If the target comparison result indicates that the object feature extraction performance of the preset ensemble model is better than that of the target ensemble model, a target object feature extraction model is generated based on the preset ensemble model. This can improve the object feature extraction performance of the target object feature extraction model and increase the distinctiveness of the features extracted by the target object feature extraction model. Moreover, the target object feature extraction model is essentially a single object feature extraction model. Compared with the deployment process of multiple models, the forward inference time of the target object feature extraction model can be greatly shortened, thereby improving the model deployment efficiency and making it suitable for a variety of complex application scenarios.

[0132] In one specific embodiment, the above method may further include:

[0133] If the target comparison results indicate that the object feature extraction performance of the target ensemble model is better than that of the preset ensemble model, the target ensemble model is used as the preset ensemble model, and the process jumps to the step of obtaining multiple models to be ensembled from multiple object feature extraction models.

[0134] Figure 3 This is a schematic diagram illustrating the model integration update process in a method for generating an object feature extraction model according to an exemplary embodiment. For example... Figure 3As shown, firstly, before the initial model integration process, the multiple object feature extraction models have not participated in the model integration process. By conducting feature extraction performance tests on the multiple object feature extraction models, the test results of the multiple object feature extraction models can be obtained. The object feature extraction model that indicates the best object feature extraction performance in the test results is used as the preset integration model. After determining the preset ensemble model, multiple models to be ensembled are obtained from multiple object feature extraction models. This means that these multiple models are the object feature extraction models excluding the one used as the preset ensemble model. By testing the feature extraction performance of these multiple models, first test results are obtained. The object feature extraction model corresponding to the target test result is selected from these models and used as the current ensemble model. Model ensemble processing is performed between the current ensemble model and the preset ensemble model to obtain the target ensemble model. The extraction performance of the target ensemble model and the preset ensemble model is compared to test whether the accuracy of the target ensemble model is improved compared to the preset ensemble model, resulting in a target comparison result. If the target comparison result indicates that the object feature extraction performance of the target ensemble model is superior to that of the preset ensemble model, the target ensemble model is adopted as the preset ensemble model, and the process is restarted. Multiple object feature extraction models are obtained from multiple models to be integrated. The model with the best object feature extraction performance among the multiple models to be integrated is taken as the current integrated model. Based on the updated preset integrated model and the updated current integrated model, model integration processing is performed to obtain a new target integrated model. The extraction performance of the new target integrated model and the updated preset integrated model is compared to determine whether the accuracy of the target integrated model is improved relative to the preset integrated model. If the target comparison result indicates that the object feature extraction performance of the target integrated model is better than that of the preset integrated model, it can be determined that the target integrated model is still the optimal model. By repeating the model integration multiple times, the object feature extraction performance of the new target integrated model obtained each time can always be higher than the best object feature extraction performance in the previous round, until the target comparison result indicates that the object feature extraction performance of the preset integrated model is better than that of the target integrated model. The preset integrated model is then taken as the target object feature extraction model.

[0135] In the above embodiments, by employing a greedy algorithm, a target ensemble model with better object feature extraction performance than the model obtained in the previous ensemble is used as a preset ensemble model. A new target ensemble model is obtained by integrating the updated preset ensemble model with the current ensemble model. Since the greedy algorithm ensures that the object feature extraction performance of the new target ensemble model obtained each time is always higher than the best object feature extraction performance in the previous round, the object feature extraction performance of the target object feature extraction model is improved, thereby increasing the distinguishability of the features extracted by the target object feature extraction model.

[0136] Based on the target object feature extraction model obtained by the above-described object feature extraction model generation method, this disclosure also provides an object recognition method, which can be used in electronic devices such as terminals and servers, and may include the following steps:

[0137] Acquire an image of the object to be identified and at least one reference object image;

[0138] The image of the object to be identified and at least one reference object image are input into the target object feature extraction model obtained by the generation method based on the above object feature extraction model to extract features, thereby obtaining the third object feature information corresponding to the image of the object to be identified and the fourth object feature information corresponding to at least one reference object image.

[0139] Based on the third and fourth object feature information, the target object in the image to be identified is identified, and the object identification result is obtained.

[0140] In one specific embodiment, the target object can refer to the object to be identified in the image of the object to be identified. Specifically, the target object can include a face or a hand, for example, the target object can be a human face.

[0141] In one specific embodiment, the third object feature information corresponding to the image of the object to be identified can characterize the object features of the target object in the image of the object to be identified. The fourth object feature information corresponding to each reference object image can characterize the object features of the object in the reference object image. The object recognition result can include a first recognition result and a second recognition result. The first recognition result can be used to indicate a reference object image that matches the target object in the image of the object to be identified, and the second recognition result can be used to indicate that at least one object in a reference object image fails to match the target object in the image of the object to be identified.

[0142] In a specific embodiment, feature comparison processing is performed based on the third object feature information and the fourth object feature information corresponding to any reference object image to obtain at least one object association information. Each of these object association information can characterize the degree of association between the object to be identified and the objects in any reference object image. Based on the at least one object association information, the object recognition result can be determined. Specifically, feature comparison processing can be performed on the third object feature information and the fourth object feature information corresponding to each reference object image to obtain the feature similarity between the third object feature information and each reference object image. Correspondingly, this feature similarity can be one of the at least one object association information. If there is object association information in the at least one object association information that is greater than a preset association information threshold, the object recognition result can be generated based on the reference object image corresponding to the object association information indicating the highest degree of association among the at least one object association information.

[0143] In one specific embodiment, the object recognition result can be applied to various object recognition scenarios. Taking a facial recognition access control system as an example, the system can store at least one reference object image. A target object feature extraction model can be pre-deployed within the system. This model extracts features from the image of the object to be recognized and at least one reference object image acquired by the system, yielding third object feature information corresponding to the image of the object to be recognized and fourth object feature information corresponding to at least one reference object image. Based on these third and fourth object feature information, object recognition is performed on the target object in the image of the object to be recognized, resulting in an object recognition result. If the object recognition result is a first recognition result, the access control device can be locked. Furthermore, based on the first recognition result, the object identification information corresponding to the reference object image corresponding to the first recognition result can be displayed. If the object recognition result is a second recognition result, the access control device can remain locked, and a display interface corresponding to the second recognition result can be displayed, indicating that the matching failed.

[0144] Figure 4 This is a block diagram illustrating an apparatus for generating an object feature extraction model according to an exemplary embodiment. Figure 4 As shown, the device may include:

[0145] The module 410 for obtaining models to be integrated can be used to obtain multiple models to be integrated from multiple object feature extraction models. The models to be integrated are object feature extraction models that have not participated in the integration process among the multiple object feature extraction models.

[0146] The feature extraction performance testing module 420 can be used to test the feature extraction performance of multiple models to be integrated, and obtain the first test results corresponding to the multiple models to be integrated; the first test results characterize the object feature extraction performance of each of the multiple models to be integrated.

[0147] The current ensemble model determination module 430 can be used to select the current ensemble model corresponding to the target test result from multiple models to be ensembled; the target test result is the test result with the best performance in extracting the indicator object features in the first test result;

[0148] The model integration module 440 can be used to perform model integration processing on the current integrated model and the preset integrated model to obtain the target integrated model. The preset integrated model is obtained by integrating at least one model whose object feature extraction performance is better than the current integrated model among multiple object feature extraction models.

[0149] The extraction performance comparison module 450 can be used to perform extraction performance comparison processing on the target ensemble model and the preset ensemble model to obtain the target comparison result.

[0150] The target object feature extraction model determination module 460 can be used to generate a target object feature extraction model based on the preset integrated model when the target comparison result indicates that the object feature extraction performance of the preset integrated model is better than that of the target integrated model.

[0151] In one specific embodiment, the apparatus for generating the above-mentioned object feature extraction model may further include:

[0152] The preset integrated model update module can be used to adopt the target integrated model as the preset integrated model when the target comparison result indicates that the object feature extraction performance of the target integrated model is better than that of the preset integrated model, and then jump to the step of obtaining multiple models to be integrated from multiple object feature extraction models.

[0153] In one specific embodiment, the above-mentioned model acquisition module 410 may include:

[0154] An integrated object feature extraction model determination module can be used to determine at least one integrated object feature extraction model among multiple object feature extraction models;

[0155] The module for determining the model to be integrated can be used to remove at least one integrated object feature extraction model from multiple object feature extraction models to obtain multiple models to be integrated.

[0156] In one specific embodiment, the apparatus for generating the above-mentioned object feature extraction model may further include:

[0157] The sample information acquisition module can be used to acquire sample object image sets, preset category information corresponding to the sample object image sets, and target category feature information corresponding to the sample object image sets.

[0158] The first information acquisition module can be used to determine the first current training image set and the first preset category information corresponding to the first current training image set from the sample object image set and the preset category information, respectively.

[0159] The first object feature information acquisition module can be used to input the first current training image set into the first preset feature extraction model to extract features and obtain the first object feature information corresponding to the first current training image set.

[0160] The first prediction category information acquisition module can be used to determine the first prediction category information corresponding to the first current training image set based on the first object feature information and the target category feature information; the first prediction category information represents the probability that any sample object image in the first current training image set belongs to any object category;

[0161] The first loss information acquisition module can be used to determine the first loss information based on the first predicted category information and the first preset category information;

[0162] The first preset feature extraction model update module can be used to update the first preset feature extraction model based on the first loss information. Based on the updated first preset feature extraction model, the training iteration steps of determining the first current training image set and the first preset category information corresponding to the first current training image set from the sample object image set and the preset category information respectively, and updating the first preset feature extraction model based on the first loss information are repeated until the first preset convergence condition is reached.

[0163] The trained feature extraction model acquisition module can be used to determine the trained feature extraction model based on the first preset feature extraction model obtained when the first preset convergence condition is met. The trained feature extraction model can be any one of multiple object feature extraction models.

[0164] In one specific embodiment, the apparatus for generating the above-mentioned object feature extraction model may further include:

[0165] The preset category feature information acquisition module can be used to acquire preset category feature information;

[0166] The second information acquisition module can be used to determine the second current training image set and the second preset category information corresponding to the second current training image set from the sample object image set and the preset category information, respectively.

[0167] The second object feature information acquisition module can be used to input the second current training image set into the second preset feature extraction model to extract features and obtain the second object feature information corresponding to the second current training image set.

[0168] The second prediction category information acquisition module can be used to determine the second prediction category information corresponding to the second current training image set based on the second object feature information and the preset category feature information; the second prediction category information represents the probability that any sample object image in the second current training image set belongs to any object category;

[0169] The second loss information acquisition module can be used to determine the second loss information based on the second predicted category information and the second preset category information;

[0170] The information and model update module can be used to update the preset category feature information and the second preset feature extraction model based on the second loss information. Based on the updated preset category feature information and the updated second preset feature extraction model, the iterative steps of determining the second current training image set and the second preset category information corresponding to the second current training image set from the sample object image set and the preset category information, respectively, and updating the preset category feature information and the second preset feature extraction model based on the second loss information are repeated until the second preset convergence condition is reached.

[0171] The target category feature information acquisition module can be used to obtain the preset category feature information obtained when the second preset convergence condition is met, and use it as the target category feature information.

[0172] In one specific embodiment, the above-mentioned performance comparison module 450 may include:

[0173] The ensemble model performance testing module can be used to test the feature extraction performance of a target ensemble model and a preset ensemble model, and obtain a second test result for the target ensemble model and a third test result for the preset ensemble model; the second test result characterizes the object feature extraction performance of the target ensemble model; the third test result characterizes the object feature extraction performance of the preset ensemble model.

[0174] The target comparison result acquisition module can be used to compare the second test result and the third test result to obtain the target comparison result.

[0175] This disclosure also provides an object recognition device, including:

[0176] The image acquisition module can be used to acquire an image of the object to be identified and at least one reference object image;

[0177] The feature extraction module can be used to input the image of the object to be identified and at least one reference object image into the target object feature extraction model for feature extraction, so as to obtain the third object feature information corresponding to the image of the object to be identified and the fourth object feature information corresponding to at least one reference object image; the target object feature extraction model is obtained based on the above object feature extraction model generation method;

[0178] The object recognition module can be used to identify target objects in an image of an object to be identified based on third and fourth object feature information, and obtain object recognition results.

[0179] Regarding the apparatus in the above embodiments, the specific manner in which each module and unit performs its operations has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0180] Figure 5 This is a block diagram illustrating an electronic device for generating a feature extraction model of a target object or identifying a target object, according to an exemplary embodiment. The electronic device may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for generating an object feature extraction model or an object recognition method.

[0181] Figure 6 This is a block diagram illustrating another electronic device for generating a feature extraction model of a target object or identifying a target object, according to an exemplary embodiment. The electronic device may be a terminal, and its internal structure diagram may be as follows: Figure 6As shown, the electronic device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for generating an object feature extraction model or an object recognition method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0182] Those skilled in the art will understand that Figure 5 or Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present disclosure and does not constitute a limitation on the electronic device to which the present disclosure is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0183] In an exemplary embodiment, an electronic device is also provided, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement a method for generating an object feature extraction model or an object recognition method as described in the embodiments of this disclosure.

[0184] In an exemplary embodiment, a computer-readable storage medium is also provided, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the object feature extraction model generation method or object recognition method of the embodiments of this disclosure.

[0185] In an exemplary embodiment, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the object feature extraction model generation method or object recognition method of the embodiments of this disclosure.

[0186] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0187] It is understood that in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0188] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0189] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. An object recognition method, characterized in that, The method includes: Acquire an image of the object to be identified and at least one reference object image; The image of the object to be identified and the at least one reference object image are input into the target object feature extraction model for feature extraction to obtain the third object feature information corresponding to the image of the object to be identified and the fourth object feature information corresponding to the at least one reference object image; Based on the third object feature information and the fourth object feature information, object recognition is performed on the target object in the image to be identified, and the object recognition result is obtained. The target object feature extraction model is obtained through the following methods: Multiple models to be integrated are obtained from multiple object feature extraction models, wherein the multiple models to be integrated are object feature extraction models that have not participated in the integration process among the multiple object feature extraction models; The feature extraction performance of the plurality of models to be integrated is tested to obtain a first test result corresponding to the plurality of models to be integrated; the first test result characterizes the object feature extraction performance of each of the plurality of models to be integrated. The current integrated model corresponding to the target test result is selected from the plurality of models to be integrated; the target test result is the test result with the best performance in extracting the indicator object features among the first test results; The current ensemble model and the preset ensemble model are subjected to model ensemble processing to obtain the target ensemble model. The preset ensemble model is obtained by ensemble of at least one of the multiple object feature extraction models whose object feature extraction performance is better than that of the current ensemble model. The target ensemble model and the preset ensemble model are subjected to performance extraction comparison processing to obtain the target comparison result; If the target comparison result indicates that the object feature extraction performance of the preset ensemble model is better than that of the target ensemble model, the target object feature extraction model is generated based on the preset ensemble model.

2. The method according to claim 1, characterized in that, The method further includes: If the target comparison result indicates that the object feature extraction performance of the target ensemble model is better than that of the preset ensemble model, the target ensemble model is used as the preset ensemble model, and the process jumps to the step of obtaining multiple models to be ensembled from multiple object feature extraction models.

3. The method according to claim 2, characterized in that, The process of obtaining multiple models to be integrated from multiple object feature extraction models includes: Identify at least one integrated object feature extraction model among the plurality of object feature extraction models; At least one integrated object feature extraction model is removed from the plurality of object feature extraction models to obtain the plurality of models to be integrated.

4. The method according to claim 1, characterized in that, The method further includes: Acquire a set of sample object images, preset category information corresponding to the set of sample object images, and target category feature information corresponding to the set of sample object images; A first current training image set and the first preset category information corresponding to the first current training image set are determined from the sample object image set and the preset category information, respectively. The first current training image set is input into the first preset feature extraction model to extract features, thereby obtaining the first object feature information corresponding to the first current training image set; Based on the first object feature information and the target category feature information, the first predicted category information corresponding to the first current training image set is determined; the first predicted category information represents the probability that any sample object image in the first current training image set belongs to any object category; Based on the first predicted category information and the first preset category information, the first loss information is determined; Based on the first loss information, the first preset feature extraction model is updated. Based on the updated first preset feature extraction model, the training iteration steps of determining the first current training image set and the first preset category information corresponding to the first current training image set from the sample object image set and the preset category information respectively, up to updating the first preset feature extraction model based on the first loss information, are repeated until the first preset convergence condition is reached. Based on the first preset feature extraction model obtained when the first preset convergence condition is met, a trained feature extraction model is determined, wherein the trained feature extraction model is any one of the plurality of object feature extraction models.

5. The method according to claim 4, characterized in that, The target category feature information is obtained using the following methods: Obtain preset category feature information; A second current training image set and the second preset category information corresponding to the second current training image set are determined from the sample object image set and the preset category information, respectively. The second current training image set is input into the second preset feature extraction model to extract features, thereby obtaining the second object feature information corresponding to the second current training image set; Based on the second object feature information and the preset category feature information, the second predicted category information corresponding to the second current training image set is determined; The second predicted category information represents the probability that any sample object image in the second current training image set belongs to any object category; Based on the second predicted category information and the second preset category information, the second loss information is determined; Based on the second loss information, the preset category feature information and the second preset feature extraction model are updated. Based on the updated preset category feature information and the updated second preset feature extraction model, the iterative steps of determining the second current training image set and the second preset category information corresponding to the second current training image set from the sample object image set and the preset category information respectively, up to the step of updating the preset category feature information and the second preset feature extraction model based on the second loss information, are repeated until the second preset convergence condition is reached. The preset category feature information obtained when the second preset convergence condition is met is used as the target category feature information.

6. The method according to claim 1, characterized in that, The step of performing performance extraction comparison processing on the target ensemble model and the preset ensemble model to obtain the target comparison result includes: The target ensemble model and the preset ensemble model are subjected to feature extraction performance tests to obtain a second test result corresponding to the target ensemble model and a third test result corresponding to the preset ensemble model; the second test result characterizes the object feature extraction performance of the target ensemble model; the third test result characterizes the object feature extraction performance of the preset ensemble model. The second test result and the third test result are compared to obtain the target comparison result.

7. An object recognition device, characterized in that, The device includes: The module for obtaining models to be integrated is used to obtain multiple models to be integrated from multiple object feature extraction models, wherein the models to be integrated are object feature extraction models that have not participated in the integration process among the multiple object feature extraction models. The feature extraction performance testing module is used to perform feature extraction performance testing on the plurality of models to be integrated, and obtain a first test result corresponding to the plurality of models to be integrated; the first test result characterizes the object feature extraction performance of each of the plurality of models to be integrated. The current ensemble model determination module is used to select the current ensemble model corresponding to the target test result from the plurality of models to be integrated; the target test result is the test result with the best performance in the first test result indicating the object feature extraction performance. The model integration module is used to perform model integration processing on the current integrated model and the preset integrated model to obtain a target integrated model. The preset integrated model is obtained by integrating at least one of the multiple object feature extraction models whose object feature extraction performance is better than the current integrated model. An extraction performance comparison module is used to perform extraction performance comparison processing on the target ensemble model and the preset ensemble model to obtain the target comparison result; The target object feature extraction model determination module is used to generate a target object feature extraction model based on the preset integrated model when the target comparison result indicates that the object feature extraction performance of the preset integrated model is better than that of the target integrated model. The image acquisition module is used to acquire an image of the object to be identified and at least one reference object image; The feature extraction module is used to input the image of the object to be identified and the at least one reference object image into the target object feature extraction model for feature extraction, so as to obtain the third object feature information corresponding to the image of the object to be identified and the fourth object feature information corresponding to the at least one reference object image; The object recognition module is used to perform object recognition on the target object in the image to be recognized based on the third object feature information and the fourth object feature information, and obtain the object recognition result.

8. The apparatus according to claim 7, characterized in that, The device further includes: The preset integrated model update module is used to, when the target comparison result indicates that the object feature extraction performance of the target integrated model is better than that of the preset integrated model, adopt the target integrated model as the preset integrated model and jump to the step of obtaining multiple models to be integrated from multiple object feature extraction models.

9. The apparatus according to claim 8, characterized in that, The module for acquiring the model to be integrated includes: An integrated object feature extraction model determination module is used to determine at least one integrated object feature extraction model among the plurality of object feature extraction models; The module for determining the model to be integrated is used to remove at least one integrated object feature extraction model from the plurality of object feature extraction models to obtain the plurality of models to be integrated.

10. The apparatus according to claim 7, characterized in that, The device further includes: The sample information acquisition module is used to acquire a sample object image set, preset category information corresponding to the sample object image set, and target category feature information corresponding to the sample object image set. The first information acquisition module is used to determine the first current training image set and the first preset category information corresponding to the first current training image set from the sample object image set and the preset category information, respectively. The first object feature information acquisition module is used to input the first current training image set into the first preset feature extraction model to extract features and obtain the first object feature information corresponding to the first current training image set. The first prediction category information acquisition module is used to determine the first prediction category information corresponding to the first current training image set based on the first object feature information and the target category feature information; the first prediction category information represents the probability that any sample object image in the first current training image set belongs to any object category; The first loss information acquisition module is used to determine the first loss information based on the first predicted category information and the first preset category information; The first preset feature extraction model update module is used to update the first preset feature extraction model based on the first loss information, and repeat the training iteration steps of determining the first current training image set and the first preset category information corresponding to the first current training image set from the sample object image set and the preset category information respectively, up to updating the first preset feature extraction model based on the first loss information, based on the updated first preset feature extraction model, until the first preset convergence condition is reached. The trained feature extraction model acquisition module is used to determine the trained feature extraction model based on the first preset feature extraction model obtained when the first preset convergence condition is met, wherein the trained feature extraction model is any one of the plurality of object feature extraction models.

11. The apparatus according to claim 10, characterized in that, The device further includes: The preset category feature information acquisition module is used to acquire preset category feature information; The second information acquisition module is used to determine the second current training image set and the second preset category information corresponding to the second current training image set from the sample object image set and the preset category information, respectively. The second object feature information acquisition module is used to input the second current training image set into the second preset feature extraction model to extract features and obtain the second object feature information corresponding to the second current training image set. The second prediction category information acquisition module is used to determine the second prediction category information corresponding to the second current training image set based on the second object feature information and the preset category feature information; the second prediction category information represents the probability that any sample object image in the second current training image set belongs to any object category; The second loss information acquisition module is used to determine the second loss information based on the second predicted category information and the second preset category information; The information and model update module is used to update the preset category feature information and the second preset feature extraction model based on the second loss information. Based on the updated preset category feature information and the updated second preset feature extraction model, the iterative steps of determining the second current training image set and the second preset category information corresponding to the second current training image set from the sample object image set and the preset category information respectively, up to updating the preset category feature information and the second preset feature extraction model based on the second loss information, are repeated until the second preset convergence condition is reached. The target category feature information acquisition module is used to take the preset category feature information obtained when the second preset convergence condition is met as the target category feature information.

12. The apparatus according to claim 7, characterized in that, The extraction performance comparison module includes: An ensemble model performance testing module is used to perform feature extraction performance tests on the target ensemble model and the preset ensemble model, and obtain a second test result corresponding to the target ensemble model and a third test result corresponding to the preset ensemble model; the second test result characterizes the object feature extraction performance of the target ensemble model; the third test result characterizes the object feature extraction performance of the preset ensemble model. The target comparison result acquisition module is used to compare the second test result and the third test result to obtain the target comparison result.

13. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to execute the executable instructions to implement the object recognition method according to any one of claims 1 to 6.

14. A non-volatile computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the object recognition method according to any one of claims 1 to 6.

15. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the object recognition method according to any one of claims 1 to 6.

Images