Target recognition method and device, electronic device and storage medium

Through segmentation network and evaluation mechanism, combined with prior knowledge in human recognition, images are segmented and classified, which solves the problem of insufficient robustness of DNNs in adversarial samples, and improves the robustness and accuracy of target recognition.

CN114612824BActive Publication Date: 2024-12-03TSINGHUA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210224960.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-09
Publication Date
2024-12-03
Estimated Expiration
2042-03-09

AI Technical Summary

Technical Problem

The existing deep neural network (DNNs)-based target recognition methods are not robust enough when facing adversarial samples, making it difficult to generalize to new attack types, and the model has a large gap in accuracy between adversarial samples and clean samples, and Prior knowledge in human identification is not fully utilized.

Method used

The image is segmented through the segmentation network, and the component feature maps of the target object are obtained, and the confidence and matching degree evaluation are performed on these feature maps. Combined with the prior knowledge in the human recognition process, the category information of the target object is determined.

Benefits of technology

The robustness and anti-interference ability of target recognition are improved, making it difficult for disturbed information to affect the overall identification process, and improving the accuracy and robustness of the identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114612824B_ABST
    Figure CN114612824B_ABST
Patent Text Reader

Abstract

The present disclosure relates to a target recognition method and device, an electronic device and a storage medium, the method comprising: performing segmentation processing on a to-be-processed image through a segmentation network to obtain a first feature map of a component of each prediction category of a target object in the to-be-processed image; evaluating the first feature map of each prediction category respectively to obtain an evaluation score of each prediction category; and determining the category information to which the target object belongs in each prediction category according to the evaluation score. According to the target recognition method of an embodiment of the present disclosure, the components of the target object can be segmented and classified in a manner similar to the process of human identification using prior knowledge, and each prediction category can be evaluated to determine the category of the target object. Since each component can be classified separately, it is difficult for the disturbance information to have a significant impact on the overall recognition process of the target object, which can improve the robustness and anti-interference ability of target recognition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a target recognition method and device, an electronic device, and a storage medium. Background Art

[0002] Object recognition refers to the use of computers to identify the types of objects in an image. With the rise of deep learning technology, the current mainstream object recognition algorithms are usually based on deep neural networks (DNNs). For example, on the ImageNet dataset, DNNs-based object recognition methods can even surpass humans in recognition accuracy. However, these method models are still far behind humans in terms of robustness. Among them, one of the most worrying things about robustness is the adversarial samples of DNNs: by adding a specific perturbation to the input image, even a tiny pixel-level perturbation that is difficult for humans to detect, DNNs will make errors that are difficult for humans to understand and greatly reduce the recognition accuracy. Among them, adversarial samples exist not only in the digital world (for example, adding a specific perturbation to the target in the image), but also in the physical world (for example, adding a specific perturbation to the target and then obtaining the target's image). This brings huge hidden dangers to the recognition algorithm based on DNNs in application. Therefore, improving the robustness of the recognition model based on DNNs has very important practical significance and application value.

[0003] In the related technology, there are many ways to improve the defense effect of DNNs against adversarial samples, including: image preprocessing, removing potential adversarial noise through methods such as denoising; loss function optimization: modifying the Softmax Cross-Entroy loss function commonly used in standard classification models to make the decision boundary of the model more obvious so as to correctly identify some adversarial samples: gradient constraint: reducing the space for adversarial samples by constraining the input gradient through supervisory signals; adversarial training: incorporating the generated adversarial samples into the training set of the DNNs model online, in order to improve the recognition effect of the deep learning model on adversarial samples.

[0004] Although the defense methods in the related art can have a certain effect on some existing adversarial sample attack methods, most methods can still be broken by adaptive attacks against the defense methods, for example, adding special perturbations in a targeted manner to make the above defense methods difficult to identify. For example, adversarial training is currently recognized as the best defense method, however, adversarial training also has the problem of not being able to generalize well to new attack types. In addition, even for neural network models after adversarial training, there is still a big gap between their accuracy on adversarial samples and their accuracy on clean samples. In addition, the current defense methods do not consider the role of human component-based recognition process and human prior knowledge in improving model robustness. Obviously, the human recognition process is difficult to be disturbed by adversarial samples. According to the conclusions of cognitive psychology, humans tend to recognize objects by components, and humans have a lot of common sense knowledge, but the current DNNs-based recognition model is usually only supervised by the category labels of objects in the picture. This will inevitably make the model learn some features that are difficult for humans to understand, resulting in errors that are difficult for humans to understand. Summary of the invention

[0005] The present disclosure proposes a target recognition method and device, an electronic device and a storage medium.

[0006] According to one aspect of the present disclosure, a target recognition method is provided, comprising: performing segmentation processing on an image to be processed through a segmentation network to obtain a first feature map of components of each predicted category of a target object in the image to be processed, wherein the components of the predicted category include, when the category of the target object is identified as the predicted category, at least one structure constituting the target object; evaluating at least one first feature map of each predicted category respectively to obtain an evaluation score of each predicted category, wherein the evaluation includes a confidence evaluation and a matching evaluation, and the evaluation score includes a confidence evaluation score and a matching evaluation score, wherein the confidence evaluation score is positively correlated with the confidence of the predicted category, and the matching evaluation score is positively correlated with the accuracy of the connection relationship of the components of the predicted category; according to the evaluation score of each predicted category, determining, in each predicted category, the category information to which the target object in the image to be processed belongs.

[0007] In a possible implementation, a segmentation process is performed on an image to be processed through a segmentation network to obtain a first feature map of a component of each predicted category of a target object in the image to be processed, including: segmenting the image to be processed through the segmentation network to obtain a second feature map of each component and a third feature map of a background area; grouping the second feature maps according to the predicted category of each component to obtain a second feature map group of each predicted category; and adding the third feature map to the second feature map group of each predicted category to obtain a first feature map of a component of each predicted category.

[0008] In a possible implementation, at least one first feature map of each prediction category is evaluated respectively to obtain an evaluation score for each prediction category, including: obtaining a maximum response feature map of the prediction category based on the at least one first feature map of the prediction category, wherein the pixel value of each pixel point of the maximum response feature map is the maximum value of the pixel value of the corresponding pixel point in the at least one first feature map; obtaining an index feature map corresponding to the maximum response feature map based on the maximum response feature map and at least one first feature map of the prediction category, wherein the pixel value of each pixel point of the index feature map is the index information of the corresponding pixel point in the maximum response feature map in the at least one first feature map; obtaining the confidence evaluation score based on the maximum response feature map and the index feature map.

[0009] In a possible implementation, at least one first feature map of each prediction category is evaluated respectively to obtain an evaluation score of each prediction category, including: performing statistics on the connection relationship of at least one maximum connected domain in the index feature map to obtain the matching evaluation score.

[0010] In a possible implementation, at least one first feature map of each prediction category is evaluated respectively to obtain an evaluation score of each prediction category, including: presetting multiple connection rules to obtain a connection rule set; counting the number of occurrences of each connection rule in a sample set including multiple sample images; obtaining the connection relationship of the components of the prediction category; and obtaining a matching evaluation score of the prediction category based on the connection relationship and the number of occurrences.

[0011] In a possible implementation, at least one first feature map of each prediction category is evaluated respectively to obtain an evaluation score of each prediction category, including: performing weighted sum processing on the confidence evaluation score and the matching evaluation score to obtain the evaluation score of the prediction category.

[0012] In a possible implementation, the method further includes: training the segmentation network using a training image having labeled information of components of the target object.

[0013] According to one aspect of the present disclosure, a target recognition device is provided, including: a segmentation module 11, used to perform segmentation processing on a to-be-processed image through a segmentation network, and obtain a first feature map of components of each predicted category of a target object in the to-be-processed image, wherein the components of the predicted category include, when the category of the target object is identified as the predicted category, at least one structure constituting the target object; an evaluation module 12, used to evaluate at least one first feature map of each predicted category respectively, and obtain an evaluation score of each predicted category, wherein the evaluation includes a confidence evaluation and a matching evaluation, and the evaluation score includes a confidence evaluation score and a matching evaluation score, wherein the confidence evaluation score is positively correlated with the confidence of the predicted category, and the matching evaluation score is positively correlated with the accuracy of the connection relationship of the components of the predicted category; a category determination module 13, used to determine, in each predicted category, the category information to which the target object in the to-be-processed image belongs according to the evaluation score of each predicted category.

[0014] In a possible implementation, the segmentation module is further used to: perform segmentation processing on the image to be processed through the segmentation network to obtain a second feature map of each component and a third feature map of the background area; group the second feature maps according to the predicted categories of the components to obtain a second feature map group of each predicted category; add the third feature map to the second feature map group of each predicted category to obtain a first feature map of the component of each predicted category.

[0015] In a possible implementation, the evaluation module is further used to: obtain a maximum response feature map of the prediction category based on at least one first feature map of the prediction category, wherein the pixel value of each pixel point of the maximum response feature map is the maximum value of the pixel value of the corresponding pixel point in the at least one first feature map; obtain an index feature map corresponding to the maximum response feature map based on the maximum response feature map and at least one first feature map of the prediction category, wherein the pixel value of each pixel point of the index feature map is the index information of the corresponding pixel point in the maximum response feature map in the at least one first feature map; obtain the confidence evaluation score based on the maximum response feature map and the index feature map.

[0016] In a possible implementation, the evaluation module is further used to: perform statistics on the connection relationship of at least one maximum connected domain in the index feature graph to obtain the matching evaluation score.

[0017] In one possible implementation, the evaluation module is further used to: preset multiple connection rules to obtain a connection rule set; count the number of occurrences of each connection rule in a sample set consisting of multiple sample images; obtain the connection relationship of the components of the predicted category; and obtain a matching evaluation score of the predicted category based on the connection relationship and the number of occurrences.

[0018] In a possible implementation, the evaluation module is further used to: perform weighted sum processing on the confidence evaluation score and the matching evaluation score to obtain the evaluation score of the predicted category.

[0019] In a possible implementation manner, the device further includes: a training module, configured to train the segmentation network using training images having labeled information of components of the target object.

[0020] According to one 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 call the instructions stored in the memory to execute the above method.

[0021] According to one aspect of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored, and the computer program instructions implement the above method when executed by a processor.

[0022] According to the target recognition method of the embodiment of the present disclosure, the components of the target object can be segmented and classified in a manner similar to the process of human recognition using prior knowledge, and each predicted category can be evaluated to determine the category of the target object. Since each component can be classified separately, it is difficult for the disturbance information to have a significant impact on the overall recognition process of the target object, which can improve the robustness and anti-interference ability of target recognition.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only and do not limit the present disclosure. Other features and aspects of the present disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used to illustrate the technical solutions of the present disclosure together with the specification.

[0025] Figure 1 A flow chart showing a target recognition method according to an embodiment of the present disclosure;

[0026] Figure 2 A schematic diagram showing an index feature map according to an embodiment of the present disclosure;

[0027] Figure 3 A schematic diagram showing a matching evaluation score according to an embodiment of the present disclosure;

[0028] Figure 4 A schematic diagram showing an application of a target recognition method according to an embodiment of the present disclosure;

[0029] Figure 5 A block diagram of a target recognition device according to an embodiment of the present disclosure is shown;

[0030] Figure 6 A block diagram of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0031] Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0032] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0033] The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can represent including any one or more elements selected from the set consisting of A, B, and C.

[0034] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present disclosure.

[0035] Figure 1 A flow chart of a target recognition method according to an embodiment of the present disclosure is shown as follows: Figure 1 As shown, the method includes:

[0036] In step S11, a segmentation process is performed on the image to be processed by a segmentation network to obtain a first feature map of components of each predicted category of the target object in the image to be processed, wherein the components of the predicted category include at least one structure constituting the target object when the category of the target object is identified as the predicted category;

[0037] In step S12, at least one first feature map of each prediction category is evaluated respectively to obtain an evaluation score of each prediction category, wherein the evaluation includes a confidence evaluation and a matching evaluation, and the evaluation score includes a confidence evaluation score and a matching evaluation score, wherein the confidence evaluation score is positively correlated with the confidence of the prediction category, and the matching evaluation score is positively correlated with the accuracy of the connection relationship of the components of the prediction category;

[0038] In step S13, based on the evaluation scores of the prediction categories, the category information to which the target object in the image to be processed belongs is determined in each prediction category.

[0039] According to the target recognition method of the embodiment of the present disclosure, the components of the target object can be segmented and classified in a manner similar to the process of human recognition using prior knowledge, and each predicted category can be evaluated to determine the category of the target object. Since each component can be classified separately, it is difficult for the disturbance information to have a significant impact on the overall recognition process of the target object, which can improve the robustness and anti-interference ability of target recognition.

[0040] In one possible implementation, the present disclosure can train a segmentation network, which is a neural network model that can segment the various components of a target object. In the related art, the training process of a neural network model that can segment or identify a target object in an image usually requires multiple images with labeled information of the target object, and each image is processed by the neural network model to obtain a segmentation result, and the error between the segmentation result and the labeled information of the target object is used to feedback and adjust the parameters of the neural network model, so that the neural network model has the function of identifying and segmenting the target object. The training of the segmentation network disclosed in the present disclosure is different from the related art. During the training process, the labeled information in the training image used for training is not the labeled information of the target object in the image, but the labeled information of the various components of the target object. The method also includes: training the segmentation network through a training image with labeled information of the components of the target object.

[0041] In the example, a training image is an image of a bird. When using it to train a neural network in the related art, the category of the bird and the area where the bird is located in the image can be annotated as a whole. When training the segmentation network of the present disclosure, the various components of the bird and the areas where they are located can be annotated, for example, the area where the bird's head is located, the area where the bird's body is located, the area where the bird's wings are located, the area where the bird's legs are located, etc.

[0042] In one possible implementation, the segmentation network can segment each component in the training image to obtain a feature map of each component, and the feature map of each component can represent the category of the component and the area where the component is located in the training image. Then, the category and area can be compared with the category and area described in the annotation information to determine the error of the segmentation network, that is, the loss function. Then, the parameters of the segmentation network can be adjusted by back propagation so that the loss function is gradually reduced. When the loss function meets the error requirements, the training is completed to obtain the trained segmentation network.

[0043] In a possible implementation, in step S11, the segmentation network can be used to perform segmentation processing on any image to be processed, and obtain feature maps of various components of the target object in the image to be processed, as well as the categories of the components in the feature maps. The category of the component may include the predicted category of the target object to which the component belongs, and the category of the component itself. For example, if the category of a feature map is "bird wings", the predicted category of the target object to which it belongs is "bird", and its own category is "wings". The present disclosure does not limit the specific category of the feature map.

[0044] In a possible implementation, the feature map is a feature map representing a component of the target object, and the feature map is obtained by the segmentation network, that is, the component in the image to be processed is segmented by the segmentation network. In the example, the feature map obtained by the segmentation network can represent the probability response to the component. For example, the segmentation network can solve the probability of whether each pixel belongs to the component. If a pixel belongs to the component, the pixel value (that is, the probability response) of the pixel is high, otherwise, the pixel value is low, and then it can be judged whether each pixel belongs to the component based on the probability response. Further, multiple pixels belonging to the component can constitute the area where the component is located.

[0045] In the example, in the feature map, the pixel value of each pixel can be expressed as the probability that the pixel belongs to the component. For example, taking the feature map of "bird's wings" as an example, the pixel value of a pixel in the feature map can be expressed as the probability that the pixel belongs to the area where the "bird's wings" are located. Therefore, in the feature map, the pixel values ​​of the pixels in the area where the "bird's wings" are located are higher, while the pixel values ​​of the pixels in other areas are lower, that is, the probability that the pixels in the area where the "bird's wings" are located belong to the area where the "bird's wings" are located is higher, for example, reaching or approaching 1, and the probability of other areas being lower, for example, reaching or approaching 0. The present disclosure does not impose any restrictions on the specific values ​​of the pixel values ​​of the pixels of the feature map.

[0046] In a possible implementation, the segmentation network can segment feature maps of multiple components of the target object and the background area. These feature maps can be grouped according to the prediction category of the target object to which they belong, and the first feature map of the component of each prediction category can be obtained. Step S11 may include: segmenting the image to be processed through the segmentation network to obtain the second feature map of each component and the third feature map of the background area; grouping the second feature map according to the prediction category of each component to obtain the second feature map group of each prediction category; adding the third feature map to the second feature map group of each prediction category to obtain the first feature map of the component of each prediction category.

[0047] In a possible implementation, as described above, the category of the feature map (i.e., the second feature map) of each component obtained by the segmentation network includes the predicted category of the target object to which the component belongs, and the category of the component itself, and can be grouped based on the predicted category of the target object. For example, in each feature map of an image to be processed including birds, the second feature map of "bird's head", the second feature map of "bird's body", and the second feature map of "bird's wings" can be grouped based on the predicted category of the target object to which they belong, which is "bird". In the segmentation process of the segmentation network, misclassification of components may also occur. For example, a branch may be misclassified as a "dog's tail", etc. Therefore, based on the predicted category of the target object to which it belongs, which is dog, the second feature map can be grouped. Generally, in the grouping of the correct predicted category, the number of second feature maps is larger, and in the grouping of the incorrect predicted category, the number of second feature maps is smaller.

[0048] In one possible implementation, the feature map obtained by the segmentation network may include (K+1)×H×W dimensions (K, H and W are all positive integers), where H×W represents the size of the feature map, K represents the number of second feature maps, that is, the number of segmented components, and (K+1) represents the total number of components and backgrounds, that is, the total number of second feature maps and third feature maps. In the above grouping process, grouping can be performed according to the predicted category of the target object to which it belongs. In the cth group, Kc (Kc is a positive integer) second feature maps may be included. For example, in the group of "birds", three second feature maps of "bird's head", "bird's body" and "bird's wings" may be included. That is, when the cth group is the group of "birds", Kc=3.

[0049] Further, the third feature map of the background can be added to all groups. For example, based on the predicted category, the second feature map can be divided into n (n is a positive integer) groups, then the third feature map can be copied n times and added to each group respectively. After adding the third feature map of the background, the feature map of each group is the first feature map of the group. The dimension of the first feature map of the cth group is (Kc+1)×H×W, where (Kc+1) is the total number of the second feature map and the third feature map in the group. For example, in the group of "birds", the feature maps of "bird's head", "bird's body", "bird's wings" and the background can be included. The four feature maps are the four first feature maps of the predicted category of "birds". Each first feature map in the predicted category can be used as the probability response of each component of the target object of the predicted category. For example, if the target object is classified as a bird, the four feature maps in the predicted category are the probability response to the bird's head, the probability response to the bird's body, the probability response to the bird's wings and the probability response to the background.

[0050] In a possible implementation, in step S12, the first feature graph of each prediction category may be evaluated to determine the evaluation score of each prediction category. In the evaluation process, the first feature graph of each prediction category (i.e., each component) may be evaluated, and the connection method of each component may be evaluated (for example, based on prior knowledge, the head of a bird should be connected to the body of a bird, the wings of a bird should be connected to the wings of a bird, etc.) to obtain an evaluation score, which is the basis for determining which prediction category is the category to which the target object belongs.

[0051] In a possible implementation, the first feature graph of each prediction category may be evaluated, for example, the confidence of the probability response of the first feature graph of each component may be evaluated, that is, the credibility of whether the classification of each component is correct may be evaluated. In the example, each first feature graph may be evaluated separately, or each first feature graph may be comprehensively evaluated to obtain the overall confidence of the prediction category.

[0052] In a possible implementation, the following focuses on the comprehensive evaluation method to obtain the overall confidence of each prediction category. Step S12 may include: obtaining a maximum response feature map of the prediction category based on at least one first feature map of the prediction category, the pixel value of each pixel point of the maximum response feature map being the maximum value of the pixel value of the corresponding pixel point in the at least one first feature map; obtaining an index feature map corresponding to the maximum response feature map based on the maximum response feature map and at least one first feature map of the prediction category, wherein the pixel value of each pixel point of the index feature map is the index information of the corresponding pixel point in the maximum response feature map in the at least one first feature map; obtaining the confidence evaluation score based on the maximum response feature map and the index feature map.

[0053] In a possible implementation, as described above, the pixel value of a pixel point in a feature map can be expressed as the probability of whether the pixel point belongs to the component, that is, if the pixel point belongs to the component, the pixel value of the pixel point is high, otherwise, the pixel value is low. Therefore, in the feature map of each component, the pixel value of the pixel point in the region where the component is located is high, and the pixel value of other regions is low. For example, in the predicted category of "bird", the pixel value of the region where the bird's head is located in the feature map of "bird's head" is high, and the pixel value of the region where the bird's wings are located in the feature map of "bird's wings" is high... Therefore, in the maximum response feature map, the pixel value of each pixel point is the maximum value of the pixel points at the same position in each first feature map. For example, in the region where the bird's head is located in the maximum response feature map, the pixel value of the region is equal to the pixel value of the pixel point in the region where the bird's head is located in the feature map of "bird's head", and in the region where the bird's wings are located in the maximum response feature map, the pixel value of the region is equal to the pixel value of the pixel point in the region where the bird's wings are located in the feature map of "bird's wings"... In the example, the maximum response feature map can be obtained by the following formula (1):

[0054] V c (i, j) = maxR c (i, j) (1)

[0055] Among them, V c is the maximum response feature map of the c-th predicted category, V c(i, j) is the pixel value of the pixel with coordinates (i, j) in the maximum response feature map, R c is at least one first feature map of the c-th predicted category, max R c (i, j) is the maximum value of the pixel value of the pixel point with coordinates (i, j) in the at least one first feature map.

[0056] In a possible implementation, an index feature map corresponding to the maximum response feature map may be determined, wherein the pixel value of each pixel in the index feature map is the index information of the corresponding pixel in the maximum response feature map, that is, which first feature map the pixel comes from.

[0057] In the example, the index feature map can be determined by the following formula (2):

[0058] T c (i, j) = arg max R c (i, j) (2)

[0059] Among them, T c is the index feature map of the c-th predicted category, T c (i, j) is the pixel value of the pixel with coordinates (i, j) in the index feature map, and arg max is the index function, which indicates the pixel value of the pixel with coordinates (i, j) in the maximum response feature map is equal to the pixel value of the pixel with coordinates (i, j) in the first feature map.

[0060] Figure 2 A schematic diagram of an index feature map according to an embodiment of the present disclosure is shown as follows: Figure 2 As shown in the figure, the pixel value of the pixel with coordinates (1, 1) is 3, indicating that the pixel value of the pixel with coordinates (1, 1) in the third first feature map is the largest, that is, in the maximum response feature map, the pixel value of the pixel with coordinates (1, 1) is equal to the pixel value of the pixel with coordinates (1, 1) in the third first feature map; similarly, the pixel value of the pixel with coordinates (1, 2) is 1, indicating that the pixel value of the pixel with coordinates (1, 2) in the first first feature map is the largest, that is, in the maximum response feature map, the pixel value of the pixel with coordinates (1, 2) is equal to the pixel value of the pixel with coordinates (1, 2) in the first first feature map.

[0061] In a possible implementation, a confidence evaluation score may be obtained according to the maximum response feature map and index feature map of each prediction category obtained above, and the confidence evaluation score may be determined according to the following formula (3):

[0062]

[0063] Among them, 1[·] is an indicative function, which can represent the index information of coordinates (i, j). If it is the third feature map of the background, the indicative function takes the value of 0, otherwise it takes the value of 1. c is a confidence evaluation score of the c-th predicted category, and the confidence evaluation score may represent an average confidence of a non-background region in at least one first feature map of the c-th predicted category.

[0064] In the example, the method of obtaining the confidence evaluation score is not limited to this, and other methods can be used to solve the confidence evaluation score. For example, the pixel value of each pixel point in the index feature map can be used as a weight, and the pixel value of each pixel point in the maximum response feature map can be weighted averaged to obtain the confidence evaluation score, etc. The present disclosure does not impose any restrictions on this.

[0065] Through the above method, the confidence evaluation score for evaluating the confidence of each prediction type can be obtained through the maximum response feature map and the index feature map, and the credibility of the prediction of each prediction type can be obtained, providing an objective data basis for determining the category of the target object.

[0066] In a possible implementation, in step S12, the evaluation score includes not only the confidence evaluation score but also the matching evaluation score, and the matching evaluation score is used to describe the connection relationship between the components of the predicted category. For example, the area where the bird's head is located is connected to the area where the bird's body is located, and the area where the bird's wings are located is connected to the area where the bird's body is located. The matching evaluation score can be used to widen the score gap between the correct predicted category and the incorrect predicted category.

[0067] In the example, if a certain predicted category is correct, the number of first feature maps in the predicted category is large, the connection relationships between the components are also large, and the accuracy of the connection relationships is also high. The wrong predicted category not only has a small number of first feature maps, but also has a small number of connection relationships between the components, and the accuracy of the connection relationships is not high. Taking the above-mentioned birds and dogs as an example, if the target object in the image to be processed is a bird, a variety of connection relationships can be detected, for example, the area where the bird's head is located is connected to the area where the bird's body is located, and the area where the bird's wings are located is connected to the area where the bird's body is located; and if objects such as branches are mistakenly identified as the tail of a dog, the tail has no connection relationship with other components of the dog, so the number of connection relationships is less than that of birds, and there is no concept of the accuracy of the connection relationship. Therefore, the matching evaluation score of birds can be made higher than that of dogs, and the possibility that the category of the target object is birds is greater than that of dogs.

[0068] In a possible implementation, step S12 may include: performing statistics on the connection relationship of at least one maximum connected domain in the index feature graph to obtain the matching evaluation score.

[0069] In a possible implementation, based on the above analysis, the matching evaluation score of each predicted category can be directly determined by the number of connection relationships, or the correctness of the connection relationship can be referred to. In the example, each predicted category can have a predefined connection relationship. For example, the predefined connection relationship of a bird may include that the area where the bird's head is located is connected to the area where the bird's body is located, the area where the bird's wings are located is connected to the area where the bird's body is located, and the area where the bird's legs are located is connected to the area where the bird's body is located. The matching evaluation score is positively correlated with the accuracy of the connection relationship of the components of the predicted category. In the example, if it is detected that the area where the bird's head is located is connected to the area where the bird's body is located, that is, it matches the predefined connection relationship (in other words, it matches the prior knowledge), then the connection relationship is correct. If the area where the bird's legs are located is connected to the area where the bird's head is located, that is, it does not match the predefined connection relationship (in other words, it does not match the prior knowledge), then the connection relationship is wrong.

[0070] In a possible implementation, in the process of performing the statistical processing, connection relationships that match predefined connection relationships can be counted. An index feature map can be used to count the connection relationships, and the pixel value of a pixel in the index feature map is the first feature map in which the pixel can obtain the maximum probability response, that is, the component to which the pixel belongs. For example, the index value of the first feature map of the bird's head area is 1, the index value of the first feature map of the bird's body area is 2, and the index value of the first feature map of the bird's wing area is 3. Then, the area with a pixel value of 1 in the index feature map can represent the bird's head area, the area with a pixel value of 2 can represent the bird's body area, and the area with a pixel value of 3 can represent the bird's wing area.

[0071] In a possible implementation, the maximum connected domain in the index feature map can be determined. For example, in the process of segmentation processing by the segmentation network, there may be some false detections of pixels, resulting in one or more isolated pixels. Determining the maximum connected domain in the index feature map can eliminate the interference of isolated pixels and obtain the area where each component is located. Further, the correct connection relationship between each maximum connected domain can be counted. For example, if the maximum connected domain with a pixel value of 1 is connected to the maximum connected domain with a pixel value of 2, the statistical number is increased by 1. If the maximum connected domain with a pixel value of 2 is connected to the maximum connected domain with a pixel value of 3, the statistical number is increased by 1. However, if the maximum connected domain with a pixel value of 1 is connected to the maximum connected domain with a pixel value of 3, the connection relationship does not match the predefined connection relationship, and the connection relationship is not counted. The statistical number finally obtained can be determined as the matching evaluation score.

[0072] Figure 3 A schematic diagram showing a matching evaluation score according to an embodiment of the present disclosure is shown as follows: Figure 3 As shown in the figure, the maximum connected domain with a pixel value of 1, the maximum connected domain with a pixel value of 2, and the maximum connected domain with a pixel value of 3 are all selected by solid-line boxes, and the area with a pixel value of 0 is the background area. In the statistical process, the maximum connected domain with a pixel value of 1 is connected to the maximum connected domain with a pixel value of 2, and the maximum connected domain with a pixel value of 2 is connected to the maximum connected domain with a pixel value of 3, and both match the predefined connection relationship. Therefore, the matching evaluation score Q can be determined. c =2.

[0073] Through the above method, the matching evaluation score can be used to further expand the distinction between the correct prediction category and the wrong prediction category, thereby improving the accuracy of the category information.

[0074] In a possible implementation, the importance of various connection relationships can also be referenced when calculating the matching evaluation score. For example, the connection between the head and body of a bird can usually be clearly presented in the image. Therefore, the connection between the two is more important for judging the bird and matches the human prior knowledge. However, the area occupied by the legs of a bird in the image is usually small and easily blocked by branches, or the legs of a bird are not easy to observe in most postures. Therefore, the importance of the connection between the legs and the body is low and does not match the features that humans often notice when identifying birds (that is, humans rarely identify birds by their legs when identifying birds). Therefore, weights can be set for various connection relationships to calculate the matching evaluation score based on the importance of the connection relationship.

[0075] In a possible implementation, step S12 may include: presetting multiple connection rules to obtain a connection rule set; counting the number of occurrences of each connection rule in a sample set consisting of multiple sample images; obtaining the connection relationship of the components of the predicted category; and obtaining a matching evaluation score of the predicted category based on the connection relationship and the number of occurrences.

[0076] In a possible implementation, the connection rule set may include a plurality of predefined connection relationships of the prediction category, i.e., connection rules. Furthermore, the number of occurrences of various connection relationships may be counted in a sample set consisting of a plurality of sample images. The more the number of occurrences, the more important the connection relationship is. For example, a sample set consisting of a plurality of sample images may include a plurality of bird images. In these bird images, the connection relationship between the bird's head and body appears in most images, while the connection relationship between the bird's legs and body appears only in a few images due to factors such as occlusion and posture. Therefore, the connection relationship between the bird's head and body appears more frequently and is more important.

[0077] In one possible implementation, the connection relationship of the components of the predicted category can be obtained. For example, the connection relationship can be counted by using the above-mentioned method of using the maximum connected domain in the index feature map, or the image to be processed itself can be directly detected to count the connection relationship of each component. Other feature maps can also be counted, for example, the connection relationship in the maximum response feature map can be counted, etc. The present disclosure does not impose any restrictions on this.

[0078] In a possible implementation, the connection relationship of the components of the predicted category can be weighted by the number of occurrences of the above connection relationship in the sample set to obtain a matching evaluation score after considering the importance factor. In the example, the matching evaluation score can be obtained by the following formula (4):

[0079]

[0080] Among them, N r Indicates the number of occurrences of the rth connection relationship in the sample set, 1 [match(r,c)] Indicates whether the rth connection relationship detected in the cth predicted category matches the predefined connection relationship, that is, whether the rth connection relationship is correct. c Represents the connection rule set.

[0081] In this way, a matching evaluation score can be obtained after considering the importance of the connection relationship, making the matching evaluation score more accurate and objective, and improving the accuracy of the category information.

[0082] In a possible implementation, after obtaining the confidence evaluation score and the matching evaluation score, the evaluation score of the predicted category can be determined based on the two. Step S12 may include: performing weighted sum processing on the confidence evaluation score and the matching evaluation score to obtain the evaluation score of the predicted category.

[0083] In this example, the evaluation score can be obtained by the following formula (5):

[0084] S c =α·P c +(1-α)·Q c (5)

[0085] Among them, S c represents the evaluation score of the cth prediction category, α represents the weight, which is also the balance coefficient between the confidence evaluation score and the matching evaluation score. The evaluation score of each prediction category can be obtained by formula (5).

[0086] In the example, the evaluation score may also be calculated in other ways, for example, directly averaging the two, or calculating the two by other formulas to obtain the evaluation score, and the present disclosure does not limit this.

[0087] In a possible implementation, in step S13, the category of the target object may be determined based on the evaluation scores of each prediction category. For example, the prediction category with the highest evaluation score may be determined as the category information of the target object in the image to be processed. * = arg max c∈{1,…,C,…} S c , that is, the predicted category corresponding to the maximum value among multiple evaluation scores.

[0088] According to the target recognition method of the embodiment of the present disclosure, the components of the target object can be segmented and classified in a manner similar to the process of human identification using prior knowledge, and each prediction category can be evaluated to determine the category of the target object. Since each component can be classified separately, it is difficult for the disturbance information to have a significant impact on the overall recognition process of the target object, which can improve the robustness and anti-interference ability of target recognition. In addition, the credibility of the prediction of each prediction type can be obtained based on the confidence evaluation score, and the distinction between the correct prediction category and the wrong prediction category can be further expanded through the matching evaluation score. When calculating the matching evaluation score, the importance of the connection relationship can be considered, so that the matching evaluation score is more accurate and objective, and the accuracy of the category information is improved.

[0089] Figure 4 A schematic diagram showing an application of the target recognition method according to an embodiment of the present disclosure is shown as follows: Figure 4As shown, the processed image (e.g., Figure 4 A bird image input into the segmentation network is segmented to obtain feature maps of each component in the image to be processed, for example, a feature map of "bird wings", a feature map of "bird head", a feature map of "bird body", and a feature map of "dog tail". The feature maps can be divided into multiple groups, that is, multiple channels, according to the predicted category of the target object described in each component. For example, the feature map of "bird wings", the feature map of "bird head" and the feature map of "bird body" are classified as the "bird" channel, and the feature map of "dog tail" is classified as the "dog" channel, and each channel includes feature maps of the predicted categories and feature maps of the background area.

[0090] In a possible implementation, the feature map of each channel can be evaluated, and the maximum response feature map of each channel can be obtained by formula (1). Taking the "bird" channel as an example, the area with the largest pixel (i.e., probability response) in each feature map is used to form the maximum response feature map of the "bird" channel, and the index feature map of each channel is obtained according to formula (2), that is, it is determined which feature map each pixel point in the maximum response feature map comes from, that is, the index information of each pixel point is obtained, and the confidence evaluation score of each channel is obtained according to formula (3). For example, the P value of the bird channel is bird .

[0091] In a possible implementation, the number of occurrences of the connection relationship of each component in the sample set can be determined based on the prior knowledge of the bird, and the connection relationship of each component of the target object in the image to be processed can be determined. The connection relationship of each component of the target object in the image to be processed can be weighted by formula (4) to obtain a matching evaluation score. For example, the Q of the bird channel is bird Similarly, the number of occurrences of the connection relationship of each component in the sample set can be determined based on the prior knowledge of the dog, and the connection relationship of each component of the target object in the image to be processed can be determined. The connection relationship of each component of the target object in the image to be processed can be weighted by formula (4) to obtain a matching evaluation score. For example, the Q of the dog channel is dog. Based on formula (4), if a certain prediction category is correct, the number of feature maps in the prediction category is large, the connection relationships between the components are also large, and the accuracy of the connection relationships is also high. The wrong prediction category not only has a small number of feature maps, but also has a small number of connection relationships between the components, and the accuracy of the connection relationships is not high. Compared with the above-mentioned birds and dogs, if the input image to be processed is an image of a bird, multiple connection relationships can be detected, for example, the area where the bird's head is located is connected to the area where the bird's body is located, and the area where the bird's wings are located is connected to the area where the bird's body is located. If an object such as a branch is mistakenly identified as a dog's tail, the tail has no connection relationship with other components of the dog, so the number of connection relationships is less than that of the bird, and there is no concept of the accuracy of the connection relationship. Therefore, the matching evaluation score of the bird can be made higher than the matching evaluation score of the dog, and the possibility that the target object is a bird is greater than that of a dog.

[0092] In a possible implementation, the evaluation score of each channel can be solved by formula (5), for example, S of the bird channel is bird , and through the arg max function, obtain the predicted category corresponding to the maximum evaluation score, and determine the predicted category as the category information of the target object in the image to be processed. In the example, the evaluation score of birds is greater than the evaluation score of dogs. Therefore, the category information can be determined to be birds, and the accuracy of category judgment is high.

[0093] In one possible implementation, the target recognition method can utilize human prior knowledge to improve the accuracy and robustness of target recognition based on the identification of each component and the evaluation of the connection relationship, so that the target recognition is more resistant to interference and more versatile. In addition, predefined connection relationships in the connection rule set, that is, connection rules, can be added to facilitate the introduction of a variety of prior knowledge to further improve the robustness of target recognition.

[0094] It can be understood that the above-mentioned various method embodiments mentioned in the present disclosure can be combined with each other to form a combined embodiment without violating the principle logic. Due to space limitations, the present disclosure will not repeat them. It can be understood by those skilled in the art that in the above-mentioned method of the specific implementation method, the specific execution order of each step should be determined according to its function and possible internal logic.

[0095] In addition, the present disclosure also provides a target identification device, an electronic device, a computer-readable storage medium, and a program, all of which can be used to implement any target identification method provided by the present disclosure. The corresponding technical solutions and descriptions are referred to the corresponding records in the method part and will not be repeated here.

[0096] Figure 5 A block diagram of a target recognition device according to an embodiment of the present disclosure is shown as follows: Figure 5 As shown, the device includes: a segmentation module 11, which is used to segment the image to be processed through a segmentation network to obtain the first feature map of the components of each predicted category of the target object in the image to be processed, wherein the components of the predicted category include, when the category of the target object is identified as the predicted category, at least one structure that constitutes the target object; an evaluation module 12, which is used to evaluate at least one first feature map of each predicted category respectively to obtain an evaluation score of each predicted category, wherein the evaluation includes a confidence evaluation and a matching evaluation, and the evaluation score includes a confidence evaluation score and a matching evaluation score, wherein the confidence evaluation score is positively correlated with the confidence of the predicted category, and the matching evaluation score is positively correlated with the accuracy of the connection relationship of the components of the predicted category; a category determination module 13, which is used to determine the category information to which the target object in the image to be processed belongs in each predicted category according to the evaluation score of each predicted category.

[0097] In a possible implementation, the segmentation module is further used to: perform segmentation processing on the image to be processed through the segmentation network to obtain a second feature map of each component and a third feature map of the background area; group the second feature maps according to the predicted categories of the components to obtain a second feature map group of each predicted category; add the third feature map to the second feature map group of each predicted category to obtain a first feature map of the component of each predicted category.

[0098] In a possible implementation, the evaluation module is further used to: obtain a maximum response feature map of the prediction category based on at least one first feature map of the prediction category, wherein the pixel value of each pixel point of the maximum response feature map is the maximum value of the pixel value of the corresponding pixel point in the at least one first feature map; obtain an index feature map corresponding to the maximum response feature map based on the maximum response feature map and at least one first feature map of the prediction category, wherein the pixel value of each pixel point of the index feature map is the index information of the corresponding pixel point in the maximum response feature map in the at least one first feature map; obtain the confidence evaluation score based on the maximum response feature map and the index feature map.

[0099] In a possible implementation, the evaluation module is further used to: perform statistics on the connection relationship of at least one maximum connected domain in the index feature graph to obtain the matching evaluation score.

[0100] In one possible implementation, the evaluation module is further used to: preset multiple connection rules to obtain a connection rule set; count the number of occurrences of each connection rule in a sample set consisting of multiple sample images; obtain the connection relationship of the components of the predicted category; and obtain a matching evaluation score of the predicted category based on the connection relationship and the number of occurrences.

[0101] In a possible implementation, the evaluation module is further used to: perform weighted sum processing on the confidence evaluation score and the matching evaluation score to obtain the evaluation score of the predicted category.

[0102] In a possible implementation manner, the device further includes: a training module, configured to train the segmentation network using training images having labeled information of components of the target object.

[0103] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0104] The embodiment of the present disclosure also provides a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above method is implemented. The computer-readable storage medium may be a non-volatile computer-readable storage medium.

[0105] An embodiment of the present disclosure further proposes an electronic device, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the above method.

[0106] The embodiments of the present disclosure also provide a computer program product, including a computer-readable code. When the computer-readable code is executed on a device, a processor in the device executes instructions for implementing the target recognition method provided in any of the above embodiments.

[0107] The embodiments of the present disclosure also provide another computer program product for storing computer-readable instructions, which, when executed, enable a computer to perform the operations of the target recognition method provided by any of the above embodiments.

[0108] The electronic device may be provided as a terminal, a server, or a device in other forms.

[0109] Figure 6 1 is a block diagram of an electronic device 1900 according to an embodiment of the present disclosure. For example, the electronic device 1900 may be provided as a terminal or a server. Figure 6, the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932 for storing instructions executable by the processing component 1922, such as an application. The application stored in the memory 1932 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above method.

[0110] The electronic device 1900 may also include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input / output (I / O) interface 1958. The electronic device 1900 may operate based on an operating system stored in the memory 1932, such as Windows Server 2000. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM or similar.

[0111] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to perform the above method.

[0112] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0113] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0114] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0115] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0116] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0117] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0118] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0119] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.

[0120] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (SDK) and the like.

[0121] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A target recognition method, characterized in that: include: Performing segmentation processing on the image to be processed by a segmentation network to obtain a first feature map of components of each predicted category of the target object in the image to be processed, wherein the components of the predicted category include at least one structure constituting the target object when the category of the target object is identified as the predicted category; Evaluate at least one first feature map of each prediction category respectively to obtain an evaluation score of each prediction category, wherein the evaluation includes a confidence evaluation and a matching evaluation, and the evaluation score includes a confidence evaluation score and a matching evaluation score, wherein the confidence evaluation score is positively correlated with the confidence of the prediction category, and the matching evaluation score is positively correlated with the accuracy of the connection relationship of the components of the prediction category; According to the evaluation scores of each prediction category, determining the category information to which the target object in the image to be processed belongs in each prediction category; The image to be processed is segmented by a segmentation network to obtain a first feature map of components of each predicted category of the target object in the image to be processed, including: The image to be processed is segmented by the segmentation network to obtain a second feature map of each component and a third feature map of the background area; According to the prediction categories of the components, the second feature maps are grouped to obtain second feature map groups of each prediction category; The third feature map is added to the second feature map group of each prediction category to obtain the first feature map of the component of each prediction category.

2. The method according to claim 1, characterized in that Evaluating at least one first feature map of each prediction category respectively to obtain an evaluation score of each prediction category includes: According to the at least one first feature map of the prediction category, a maximum response feature map of the prediction category is obtained, wherein the pixel value of each pixel point of the maximum response feature map is the maximum value of the pixel value of the corresponding pixel point in the at least one first feature map; According to the maximum response feature map and at least one first feature map of the predicted category, an index feature map corresponding to the maximum response feature map is obtained, wherein the pixel value of each pixel point of the index feature map is the index information of the corresponding pixel point in the maximum response feature map in the at least one first feature map; The confidence evaluation score is obtained according to the maximum response feature map and the index feature map.

3. The method according to claim 2, characterized in that Evaluating at least one first feature map of each prediction category respectively to obtain an evaluation score of each prediction category includes: The connection relationship of at least one largest connected domain in the index feature graph is counted to obtain the matching evaluation score.

4. The method according to claim 1, characterized in that Evaluating at least one first feature map of each prediction category respectively to obtain an evaluation score of each prediction category includes: Preset multiple connection rules and obtain connection rule sets; In a sample set consisting of a plurality of sample images, counting the number of occurrences of each of the connection rules; Obtaining a connection relationship between components of the predicted category; A matching evaluation score of the predicted category is obtained according to the connection relationship and the number of occurrences.

5. The method according to claim 1, characterized in that Evaluating at least one first feature map of each prediction category respectively to obtain an evaluation score of each prediction category includes: The confidence evaluation score and the matching evaluation score are weighted and summed to obtain the evaluation score of the predicted category.

6. The method according to claim 1, characterized in that The method further comprises: The segmentation network is trained by using training images with labeled information of components of the target object.

7. A target recognition device, characterized in that: include: A segmentation module, configured to perform segmentation processing on the image to be processed through a segmentation network to obtain a first feature map of components of each predicted category of the target object in the image to be processed, wherein the components of the predicted category include, when the category of the target object is identified as the predicted category, at least one structure constituting the target object; An evaluation module, used to evaluate at least one first feature map of each prediction category respectively to obtain an evaluation score of each prediction category, wherein the evaluation includes a confidence evaluation and a matching evaluation, and the evaluation score includes a confidence evaluation score and a matching evaluation score, wherein the confidence evaluation score is positively correlated with the confidence of the prediction category, and the matching evaluation score is positively correlated with the accuracy of the connection relationship of the components of the prediction category; A category determination module, used to determine, in each predicted category, the category information to which the target object in the image to be processed belongs according to the evaluation score of each predicted category; The segmentation module is further used for: The image to be processed is segmented by the segmentation network to obtain a second feature map of each component and a third feature map of the background area; According to the prediction categories of the components, the second feature maps are grouped to obtain second feature map groups of each prediction category; The third feature map is added to the second feature map group of each prediction category to obtain the first feature map of the component of each prediction category.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.