Object attribute identification method and device, electronic equipment and storage medium

Abstract

The invention provides an object attribute recognition method and device, electronic equipment and a storage medium, and is applied to the technical field of neural networks. The to-be-recognized image comprises a to-be-recognized object; inputting the to-be-recognized image into a preset attribute recognition network for attribute recognition processing, and determining at least one attribute of the to-be-recognized object; wherein the attribute recognition network is determined by a candidate neural network structure with the optimal performance, which is obtained by searching candidate neural network structures in a preset search space, and the preset search space comprises at least one candidate neural network structure; the network structures of all the candidate neural network structures in the preset search space are not completely the same. By adopting the technical scheme of the invention, the labor cost and the time cost for determining the attribute identification network can be reduced.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to an object attribute recognition method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the development of artificial intelligence technology, object attribute recognition (taking human attribute recognition as an example) has become an important research field, with wide applications in security, smart retail, and human-computer interaction. However, due to the diversity and complexity of object attribute recognition, training a separate network model for each attribute is impractical in real-world deployments. Therefore, designing an efficient and excellent neural network is a challenging problem that needs to be solved in the field of object attribute recognition.

[0003] In most related technologies, network structures for object recognition are designed manually based on experience. Then, the parameters of the designed network structure are manually tuned and optimized to obtain the optimized network structure, which is then used for object attribute recognition.

[0004] However, the above technologies are time-consuming and have high labor costs. Summary of the Invention

[0005] This invention provides an object attribute identification method, device, electronic device, and storage medium to solve the defects of the prior art, which is time-consuming and labor-intensive due to the manual design of network structure and manual adjustment and optimization of network structure. It aims to reduce labor and time costs by automatically determining the attribute identification network.

[0006] In a first aspect, the present invention provides an object attribute identification method, comprising the following steps: Obtain the image to be identified; the image to be identified includes the object to be identified; The image to be identified is input into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be identified; The aforementioned attribute recognition network is determined by searching for the candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

[0007] According to an object attribute recognition method provided by the present invention, the determination method of the above-mentioned attribute recognition network includes: Acquire multiple training images; each training image includes sample objects with labeled attributes, and the multiple training images constitute the training set, validation set, and test set of the sample objects; Based on multiple training images, gradient descent and progressive search strategies are used to train all candidate neural network structures in the preset search space to determine the backbone network. Based on multiple training images, gradient descent is used to train all candidate neural network structures in a preset search space to determine at least one branch network; each branch network is used to identify one attribute of the sample object. Determine the attribute identification network based on the backbone network and at least one branch network.

[0008] According to the object attribute recognition method provided by the present invention, the above-mentioned method trains all candidate neural network structures in a preset search space based on multiple training images using gradient descent and a progressive search strategy to determine the backbone network, including: All candidate neural network structures in the preset search space are connected to nodes as a network layer, and the initial first neural network is determined according to the network layer with the first preset number of layers. The training operation includes: training an initial first neural network using gradient descent based on multiple training images to obtain a converged first neural network; the first neural network includes trained candidate neural network structures and first weights corresponding to each candidate neural network structure. Perform a progressive search operation, which includes: based on the first weight of each candidate neural network structure, using a progressive search strategy to delete a first preset number of candidate neural network structures in the first neural network, and increasing the number of layers in the first neural network to obtain an intermediate first neural network; The intermediate first neural network is used as the new initial first neural network. The above training operation and progressive search operation are performed iteratively until the preset iteration cutoff condition is met. The backbone network is obtained based on the intermediate first neural network when the iteration cutoff condition is met.

[0009] According to the object attribute recognition method provided by the present invention, the above-mentioned obtaining the backbone network based on the intermediate first neural network when the iteration cutoff condition is met includes: If the number of skip connections in the intermediate first neural network that meets the iteration cutoff condition exceeds the second preset number, then the skip connection with the smallest weight in the intermediate first neural network is obtained; the above skip connection is a candidate neural network structure. Replace the jump connection with the smallest weight with the target neural network structure; the target neural network structure is the candidate neural network structure with the second largest weight in the intermediate first neural network.

[0010] According to the object attribute recognition method provided by the present invention, the above-mentioned method trains all candidate neural network structures in a preset search space using gradient descent based on multiple training images to determine at least one branch network, including: All candidate neural network structures in the preset search space are connected to nodes as a network layer, and the initial branch network is determined according to the network layer with the second preset number of layers. Based on the number of attributes of the sample objects to be identified, at least one initial branch network is constructed, and the backbone network and at least one initial branch network are combined to determine the initial second neural network; Based on multiple training images, an initial second neural network is trained using gradient descent to obtain a converged second neural network; the second neural network includes at least one converged branch network.

[0011] According to the object attribute recognition method provided by the present invention, the above-mentioned method trains an initial second neural network using gradient descent based on multiple training images to obtain a converged second neural network, comprising: Multiple training images are input into the initial second neural network for attribute recognition processing to determine at least two predicted attributes corresponding to the sample objects in each training image. For each training image, calculate the loss between each predicted attribute and the corresponding labeled attribute of the sample object in the training image, and sum the losses by weight to determine the loss value corresponding to the training image. Based on the loss and value of each training image, the initial second neural network is trained using gradient descent to obtain a converged second neural network.

[0012] According to an object attribute identification method provided by the present invention, the above-mentioned determination of the attribute identification network based on the backbone network and at least one branch network includes: The backbone network and at least one branch network are combined to determine the initial attribute recognition network; The training and validation sets of the sample objects are combined into a new training set, and the initial attribute recognition network is trained using the new training set to determine the attribute recognition network.

[0013] Secondly, the present invention also provides an object attribute recognition device, comprising the following modules: The acquisition module is used to acquire the image to be recognized; the image to be recognized includes the object to be recognized. The attribute recognition module is used to input the image to be recognized into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be recognized. The aforementioned attribute recognition network is determined by searching for the candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

[0014] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the object attribute recognition method as described above.

[0015] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the object attribute identification method as described above.

[0016] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the object attribute recognition method as described above.

[0017] The object attribute recognition method, apparatus, electronic device, and storage medium provided by this invention acquire an image of the object to be recognized, and input the image into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object. The attribute recognition network is determined by searching for at least one candidate neural network structure in a preset search space using gradient descent and a progressive search strategy, resulting in the optimal candidate neural network structure. The preset search space includes at least one candidate neural network structure with different network structures. In this method, since the candidate neural network structures in the preset search space can be automatically searched using gradient descent and a progressive search strategy to obtain the optimal neural network structure and determine the attribute recognition network accordingly, manual design and parameter tuning / optimization of the network structure are eliminated. This reduces the workload of manual network design and parameter tuning / optimization, thus lowering labor and time costs. Furthermore, the automated search of network structures using gradient descent and a progressive search strategy improves the design efficiency and accuracy of the attribute recognition network. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 It is a diagram illustrating some attributes of the human body.

[0020] Figure 2 This is one of the flowcharts illustrating the object attribute recognition method provided in this embodiment of the invention.

[0021] Figure 3 This is the second flowchart of the object attribute recognition method provided in the embodiments of the present invention.

[0022] Figure 4 This is a flowchart illustrating the progressive search network structure provided in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the architecture of the attribute recognition network provided in an embodiment of the present invention.

[0024] Figure 6 This is a schematic diagram of the overall process of the attribute recognition network provided in the embodiments of the present invention.

[0025] Figure 7 This is a schematic diagram of the object attribute recognition device provided in an embodiment of the present invention.

[0026] Figure 8 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the embodiments of the present invention.

[0028] The following will first explain the proper nouns that may be involved in the embodiments of the present invention.

[0029] NAS: Neural Architecture Search, refers to a method that uses automated search to find the optimal neural network structure in a given search space (such as different layers or structures in a convolutional neural network) to improve model performance.

[0030] Backbone: Originally meaning pillar, foundation, or backbone, in neural networks, especially in the field of CV (Computer Vision), it generally refers to the basic network part that extracts features from images. This part can be regarded as the foundation of CV tasks.

[0031] Skip-connection: A broad concept that refers to a direct connection between two locations within a network model, also known as a skip connection.

[0032] The following describes the relevant technical background of the embodiments of the present invention.

[0033] Taking human attribute recognition as an example, human attribute recognition refers to the analysis of attributes of the human body in images or videos, such as gender, age, hairstyle, and clothing, using computer vision and machine learning technologies. These technologies have wide applications in security, smart retail, and human-computer interaction. However, due to the diversity and complexity of human attribute recognition, such as... Figure 1 The diagram shown illustrates human attributes, but only a portion of them are represented. A single pedestrian image typically requires analysis of nearly ten different attributes. Therefore, in practical applications, training a separate model for each attribute is impractical. Designing an efficient and superior neural network is a critical challenge in the field of human attribute recognition. Furthermore, the three key elements of deep learning-based algorithms or solutions are algorithms, computing power, and data. For a mature solution to be truly implemented, it needs to be deployed on chips with limited performance. Therefore, improving product effectiveness and performance through performance optimization at the model algorithm level has become a common and promising area of ​​research.

[0034] Given the above, designing a reasonable and efficient neural network quickly within limited performance constraints becomes particularly important. However, designing and optimizing deep learning models is not an easy task; it requires significant expertise, experience, and time. Furthermore, different tasks and datasets may necessitate different model structures and parameter settings. Given the complexity and diversity of human attributes, relying solely on manual design of network structures is unreasonable, extremely time-consuming, and inefficient.

[0035] Currently, in the field of automatically designing neural network structures, one technique proposes a greedy search-based method for finding the optimal neural network structure. This can be used for multi-attribute image prediction, improving the efficiency and performance of search algorithms and automating the machine learning process. However, this invention only provides one search method and does not specifically design multi-branch search schemes. For characteristics such as the diversity of human attributes, multiple searches are required, making this method unsuitable for solving human attribute-related problems. Secondly, this invention designs the number of network layers from the outset, without considering network depth and matching degree. Another technique utilizes multi-attribute convolutional neural network models to simultaneously identify multiple human attributes, improving the recognition efficiency. However, this method requires manual parameter tuning and optimization when training the model for different human attributes. Designing the network structure entirely manually requires a significant amount of professional knowledge and experience, which wastes human resources and time to some extent.

[0036] Based on this, embodiments of the present invention provide an object attribute identification method, apparatus, electronic device, and storage medium, which can solve the above-mentioned technical problems.

[0037] The following is combined with Figures 2-6 This invention describes an object attribute recognition method according to an embodiment of the present invention.

[0038] It should be noted that the execution subject of the embodiments of the present invention may be an object attribute recognition device, an electronic device including an object attribute recognition device, or an object attribute recognition system including an object attribute recognition device or an electronic device, etc. The following embodiments use an electronic device as the execution subject to describe the object attribute recognition method of the embodiments of the present invention.

[0039] Figure 2 This is one of the flowcharts illustrating the object attribute recognition method provided by the present invention, such as... Figure 2 As shown, the method includes the following steps: S102, Obtain the image to be identified; the image to be identified includes the object to be identified.

[0040] In this step, the object to be identified can be a human body, an animal body, a plant body, an object, or any other object whose attributes need to be identified; there are no specific limitations here. The attributes here can be attributes that the object itself possesses or carries.

[0041] Specifically, an image to be identified can be obtained by capturing an image of the object to be identified; or a pre-captured image of the object to be identified can be stored in the cloud or local storage, and the image can be retrieved directly from the cloud or local storage when needed; or other acquisition methods can be used, which are not specifically limited here.

[0042] S104, the image to be identified is input into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be identified; wherein, the attribute recognition network is determined by searching for the candidate neural network structure in the preset search space using gradient descent and progressive search strategies to obtain the candidate neural network structure with the best performance, and the preset search space includes at least one candidate neural network structure.

[0043] In this embodiment of the invention, a preset search space can be pre-constructed. This preset search space may include one or more neural network structures (all referred to as candidate neural network structures), generally including multiple candidate neural network structures, and these multiple candidate neural network structures are not all the same. It is understood that the preset search space may include a set of various possible network architectures, such as 3×3 convolution, 1×3 convolution, residual structures, etc.; the preset search space may include different numbers of network layers, number of channels, operators (representing the type of operation, direction of operation, etc.), etc. Optionally, the above-mentioned candidate neural network structures may also be referred to as candidate operations.

[0044] It should be noted that each candidate neural network structure in the above-mentioned preset search space may include the specific structure of the neural network (such as how many convolutional layers it includes, and whether it uses 3×3 convolution) and the processing method between the specific structures (such as the connection method between each convolutional layer and the function it performs).

[0045] After constructing the preset search space, a search strategy can be designed to search for candidate network structures within the preset search space. This search for candidate network structures is called Neural Architecture Search (NAS). Generally, various search strategies can be used for neural architecture search, such as random search, evolutionary algorithms, and reinforcement learning. These search strategies find the optimal solution by evaluating and comparing different network structures. During the search process, optimization techniques such as pruning and parameter sharing can be incorporated to reduce the size of the search space. In this embodiment, a progressive search strategy is used when searching the preset search space, combined with gradient descent (also known as a differentiable approach) to improve search efficiency.

[0046] The aforementioned incremental search strategy refers to first searching all candidate neural network structures in a preset search space, then searching for the best-performing candidate neural network structures within that space based on the search results, and so on, continuously reducing the number of candidate neural network structures searched to approximate the optimal candidate neural network structure. This incremental search process can be combined with gradient descent, where gradient descent involves differentiating the loss in the process of finding the optimal candidate neural network structure at each search step to quickly determine the optimal candidate neural network structure.

[0047] After identifying the optimal candidate neural network structure in the preset search space, this optimal candidate neural network structure can be directly used as the attribute recognition network. Alternatively, it can be combined with other network structures to form an attribute recognition network. It can also be obtained by retraining the optimal candidate neural network structure, or through other methods. No specific limitations are imposed here. Furthermore, the optimal candidate neural network structure can be the one that best combines the performance of the algorithm, computing power, and data.

[0048] It is understood that the aforementioned attribute recognition network is a neural network. This network can include a backbone network and at least one branch network. During the search of this neural network structure, the backbone network and at least one branch network can be directly searched, or they can be searched separately. The backbone network is used to identify the basic attribute features of the object to be recognized in the image (e.g., identifying the target). The backbone network, along with each branch network, can be used to identify specific attributes of the object. Each branch network (including combinations of branch networks and the backbone network) is used to identify one attribute of the object (e.g., different branch networks can be set to identify different attributes such as gender or clothing style).

[0049] After obtaining the aforementioned attribute recognition network, the image to be recognized can be directly or after preprocessing and input into the attribute recognition network to identify the attributes on the object to be recognized and obtain one or more identified attributes.

[0050] In this embodiment, an image of the object to be identified is acquired, and the image is input into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object. This attribute recognition network is determined by searching for at least one candidate neural network structure in a preset search space using gradient descent and a progressive search strategy, finding the optimal candidate neural network structure. The preset search space includes at least one candidate neural network structure with different network structures. In this method, since the candidate neural network structures in the preset search space can be automatically searched using gradient descent and a progressive search strategy to obtain the optimal neural network structure and determine the attribute recognition network accordingly, manual design and parameter tuning / optimization of the network structure are eliminated. This reduces the workload of manual network design and parameter tuning / optimization, thus lowering labor and time costs. Furthermore, the automated search of network structures using gradient descent and a progressive search strategy improves the design efficiency and accuracy of the attribute recognition network.

[0051] The following examples illustrate the specific determination process of the attribute recognition network.

[0052] In some embodiments, see Figure 3 The second flowchart illustrating the object attribute recognition method shows that the determination of the aforementioned attribute recognition network may include the following steps: S202, acquire multiple training images; each training image includes sample objects with labeled attributes, and the multiple training images constitute the training set, validation set and test set of the sample objects.

[0053] This process involves pre-collecting multiple images of the same sample object and labeling the attributes of the sample object in each image to obtain images of the sample object with labeled attributes, which are then denoted as training images. This results in multiple training images. The labeled attributes of the sample objects in each training image can be a single attribute or multiple attributes; that is, training images can include single-label and / or multi-label training images.

[0054] After obtaining multiple training images with single and / or multiple labels, these multiple training images can be divided into training set, validation set and test set according to a certain ratio, such as 6:2:2.

[0055] In addition, after collecting multiple images of the same sample object, these multiple images can be preprocessed (such as cleaning and classification) before attribute labeling.

[0056] S204. Based on multiple training images, the system trains all candidate neural network structures in the preset search space using gradient descent and a progressive search strategy to determine the backbone network.

[0057] As mentioned in the above embodiment, the attribute recognition network may include a backbone network and at least one branch network. The neural network structures of the backbone network and the branch network are different. Therefore, the backbone network and the branch network can be searched separately in the preset search space. In this step, after obtaining multiple training images, these multiple training images can be used to first search for the neural network structure of the backbone network in the preset search space.

[0058] As an optional embodiment, searching the neural network structure of the backbone network may include the following steps: Step A1: Connect all candidate neural network structures in the preset search space to nodes as a network layer, and determine the initial first neural network based on the first preset number of network layers.

[0059] Among them, see Figure 4 The flowchart of the progressive search network structure shown can be a feature map after being processed by different network structures / operations. The feature map can be a feature map of the attributes of the object to be identified, such as nodes 0, 1, 2, 3, etc. in the figure.

[0060] Specifically, a corresponding number of nodes can be set according to the attribute recognition needs, and then all candidate neural network structures in the preset search space are connected to the nodes. Here, the connection refers to the connection between each node, including the connection between each node and all the nodes in front of it, that is, each node is connected to all the nodes in front of it.

[0061] Then, the network structure that connects all the candidate neural network structures between each node can be used as a network layer. The network layer with the first preset number of layers can be set according to the actual situation of attribute recognition requirements. Then, the network layers with the first preset number of layers can be connected according to the set structure. At the same time, the corresponding functions can be set for each network layer. After setting, the constructed neural network can be obtained, which is called the initial first neural network.

[0062] Each network layer here can include a Normal Cell and a Reduction Cell, as shown in the diagram. The Normal Cell ensures that the feature map remains the same size after processing by the candidate neural network structure, while the Reduction Cell halves the size of the feature map after processing by the candidate neural network structure. The feature map of each node is obtained by processing the feature maps of all its forward nodes through the candidate neural network structure, and can be represented by the following formula: (1).

[0063] in, x Represents the feature map; o This represents a candidate neural network structure / operation within a predefined search space; i , j This indicates the sequence number of the feature map / node; the specific size can be set according to the actual situation.

[0064] To ensure the continuity of the predefined search space, a softmax operation (e.g., normalization) can be performed on all candidate neural network structures, as shown in the following formula: (2).

[0065] Among them, uppercase O Indicates the preset search space; o This represents a candidate neural network structure / operation within a predefined search space; o’ This represents all candidate neural network structures / operations in the preset search space; x Representing feature maps, o ( x ) indicates that it acts on x A candidate neural network structure / operation for this feature map; This represents the result of applying a softmax operation to the feature map of a candidate neural network structure / operation on a node; a set of nodes that perform mixed weight operations on the candidate neural network structure / operation can be represented as a vector. .

[0066] After the network structure search is completed, replacing each hybrid candidate neural network structure / operation with the most probable candidate neural network structure / operation will yield a new discrete structure, as shown in the following formula: (3).

[0067] The above operations can transform the preset search space from a discrete space into a continuous space, so that gradient descent can be used to search for the optimal candidate network structure.

[0068] Furthermore, the aforementioned first preset number of layers could be, for example, 5 layers. The total number of candidate neural network structures in the aforementioned preset search space is 5, for example, there are 5 connecting lines between each node in the diagram, and each connecting line represents a candidate neural network structure.

[0069] Step A2, perform the training operation, which includes: training the initial first neural network using gradient descent based on multiple training images to obtain a converged first neural network; the first neural network includes the trained candidate neural network structure and the first weight corresponding to each candidate neural network structure.

[0070] In this step, when constructing the initial first neural network, a weight can be assigned to each candidate neural network structure. This weight is a structural parameter of the corresponding candidate neural network structure. This can represent the performance level of the corresponding candidate neural network structure. After constructing the initial first neural network, the weights of each candidate neural network structure in the initial first neural network can be initialized, that is, the architecture parameters are randomly initialized for each candidate neural network structure / operation. k = 1, 2, 3, 4, 5...

[0071] The goal of the search is to learn the aforementioned structural parameters. and network weight w (The network refers to the initial first neural network, the first neural network, and subsequent intermediate first neural networks.) Assuming we use... L train and L val Let these represent the losses for the training and validation sets, respectively. These two losses determine not only the structural parameters... And it also determines the network weight. w The goal of the search is to find This results in the loss of the validation set. Minimum, of which The loss is obtained by minimizing the training loss, and the specific calculation method is shown in the following formula: (4).

[0072] Here, st indicates subordination, i.e., constraint.

[0073] As can be seen, the above formula is a two-layer nested loss calculation formula. This two-layer nested loss calculation can search for the optimal candidate neural network structure in the preset search space. However, the optimization process of this method is time-consuming and requires significant resources. Therefore, this embodiment of the invention proposes a gradient descent two-layer approximation optimization algorithm to solve for the optimal candidate neural network structure, that is, to transform the above formula as shown in the following formula: (5).

[0074] in, wThis represents the current weights of the initial first neural network; This represents the learning rate of the initial optimization step within the first neural network.

[0075] In the embodiments of the present invention, To approximate This approach achieves the effect of single-step optimization through a two-layer approximation optimization algorithm using gradient descent, thereby improving the efficiency of network structure search and reducing resource consumption.

[0076] Specifically, after setting up the search algorithm (a two-layer approximate optimization algorithm of gradient descent) in the initial first neural network, each training image in the training set can be input into the initial first neural network for processing to obtain the processing results of each training image in the training set. The loss of the training set can then be calculated using the processing results of each training image. L train The same method is used to input each training image in the validation set into the initial first neural network for processing, obtaining the processing results of each training image in the validation set, and then calculating the loss of the validation set based on the processing results of each training image. L val Then, the gradient descent method described in formula (5) above is used to optimize the structural parameters of each candidate neural network structure / operation. This continues until the initial first neural network converges, thus obtaining a converged first neural network.

[0077] In addition, the loss function used when training the initial first neural network can be the cross-entropy loss function, such as the following formula (6): (6).

[0078] in, y Indicates the sample label value; L represents the predicted value of the sample; L represents the loss value.

[0079] After obtaining the trained first neural network, a set of structural parameters can be obtained, including the structural parameters of each candidate neural network structure / operation in the first neural network. Here, the structural parameters of each candidate neural network structure can be denoted as the first weight, for example... Figure 4 The numbers next to each connection line / candidate neural network structure in the enlarged diagram above the four nodes 0, 1, 2, and 3 represent the first weights of each candidate neural network structure in the example.

[0080] It is understandable that the above steps for training the initial first neural network can be... Figure 4 The Initial Stage.

[0081] Step A3: Perform a progressive search operation, which includes: based on the first weight of each candidate neural network structure, using a progressive search strategy to delete a first preset number of candidate neural network structures in the first neural network, and increasing the number of layers in the first neural network to obtain an intermediate first neural network.

[0082] In this step, after optimizing the initial first neural network and its structural parameters and network weights in step A2 above to obtain the first neural network, the first weights of each candidate neural network structure in the first neural network structure can be sorted (either from largest to smallest or smallest to largest) to obtain the sorting result. Then, based on the sorting result, a first preset number of candidate neural network structures with smaller first weights can be deleted. The size of this first preset number can be set according to the actual situation, for example, it can be 1, 2, 3, etc. Figure 4 The candidate neural network structure with the smallest weight is removed, reducing the number of candidate neural network structures in the initial first neural network from 5 to 4.

[0083] After deleting some candidate neural network structures from the first neural network, the number of layers in the first neural network can be increased. The specific number of layers to add can be set according to the actual situation, for example, in... Figure 4 Six new network layers are added, meaning three more Normal Cell layers are added to each of the three Normal Cell layers in the initial first neural network, resulting in a total of 11 network layers. Of course, Reduction Cell layers can also be added in practice; this is not a specific limitation. Furthermore, it should be noted that the structure of each added network layer is the same as the structure of the network layers after deleting some candidate neural network structures.

[0084] After deleting some candidate neural network structures from the first neural network and adding network layers, the resulting new first neural network can be connected to the intermediate first neural network.

[0085] It is understandable that the above method of deleting some candidate neural network structures in the first neural network and adding network layers is an incremental search strategy, which gradually approaches the candidate neural network structure with the best performance.

[0086] It is understandable that the above-described progressive steps to obtain the intermediate first neural network can be... Figure 4 The intermediate stage in the process.

[0087] Step A4: Use the intermediate first neural network as the new initial first neural network, iteratively execute the above training operation until the preset iteration cutoff condition is met, and obtain the backbone network based on the first neural network that meets the iteration cutoff condition.

[0088] In this step, after obtaining the intermediate first neural network, it can be used as the new initial first neural network. The process returns to iteratively executing step A2, "Training the initial first neural network using gradient descent based on multiple training images to obtain a converged first neural network." This initializes the structural parameters of the new initial first neural network, and after training, a new set of structural parameters is obtained, representing the new first weights for each candidate neural network structure in the new first neural network. Then, step A3 is executed again, using a progressive search strategy based on the new first weights to delete a predetermined number of candidate neural network structures and increase the number of layers in the new first neural network, thus obtaining the new intermediate first neural network. This predetermined number can be the same as or different from the aforementioned first preset number. Figure 4 In this process, two candidate neural network structures with smaller first weights can be removed. The number of added network layers can be the same as or different from the number added in A3 above, for example, in... Figure 4 This iteration adds 6 more Normal Cell network layers.

[0089] After obtaining the new intermediate first neural network, it can be determined whether the preset iteration cutoff condition is met. If it is met, the candidate neural network structure with the largest first weight in the latest intermediate first neural network is taken as the candidate neural network structure with the best performance, and the candidate neural network structure with the best performance is taken as the backbone network of the attribute recognition network.

[0090] Optionally, the above iteration cutoff condition may include any of the following: the current iteration number reaches a preset iteration number, or the number of remaining candidate neural network structures in the latest intermediate first neural network structure reaches a preset number.

[0091] For example, as mentioned above Figure 4 In the final stage of the iteration, the two candidate neural network structures with smaller first weights in the previous intermediate first neural network can be deleted. Finally, two candidate neural network structures remain in the new intermediate first neural network, and the candidate neural network structure with larger first weight is selected as the candidate neural network structure with the best performance.

[0092] S206, Based on multiple training images, train all candidate neural network structures in the preset search space using gradient descent to determine at least one branch network; each branch network is used to identify one attribute of the sample object.

[0093] In this step, after determining the backbone network, its structure can be fixed. Based on the training and validation sets, gradient descent is used to search for the optimal candidate neural network structure in the preset search space, which becomes a branch network of the attribute recognition network. If there are multiple attributes to be recognized, there can be multiple branch networks. These branch networks can be determined in this way, with each branch network used to recognize one attribute of a sample object in the training image.

[0094] S208, Determine the attribute identification network based on the backbone network and at least one branch network.

[0095] In this step, after determining the backbone and branch networks of the attribute recognition network as described above, see [link to relevant documentation]. Figure 5 The schematic diagram of the attribute recognition network architecture, shown below, uses two branch networks (Cell1 and Cell2) as an example. The backbone is the optimal performing backbone network searched in S204 above, and the "?" in Cell1 and Cell2 indicates the optimal performing branch network searched in S206 above. The backbone network and each branch network can then be connected together, and each branch network can be connected to other corresponding network structures, such as GAP average pooling layers, FC fully connected layers, and Softmax normalization layers, to collectively form the attribute recognition network.

[0096] In this embodiment, using multiple training images labeled with attributes, a backbone network is searched in a preset search space using gradient descent and a progressive search strategy. Branch networks are also searched in the preset search space using gradient descent. The attribute recognition network is then determined based on the backbone and branch networks. This gradient descent and progressive search strategy reduces search time, improves network efficiency, and better utilizes resources. Furthermore, during the backbone network search, candidate neural network structures with smaller weights are gradually removed, and the number of network layers is gradually increased, ensuring that the final determined backbone network structure is optimal. This progressive search strategy, compared to simple unit search followed by stacking, results in a backbone network structure that better reflects reality, leading to superior performance of the final attribute recognition network. Moreover, this progressive search strategy is generally more efficient than a one-time global search because it utilizes information obtained in previous search phases, avoiding repeated exploration of ineffective architectures already tried, thereby improving network search efficiency.

[0097] During the network structure search process described above, it can be observed that neural networks tend to favor skip connections without learning parameters in order to accelerate convergence. However, an excessive number of skip connections can lead to overfitting and poor performance in the final attribute recognition. Therefore, in the final determined backbone network structure, it is necessary to minimize the number of skip connections to improve the recognition performance of the final attribute recognition network. The following examples illustrate this process.

[0098] In some embodiments, step A4 above, "obtaining the backbone network based on the intermediate first neural network that satisfies the iteration cutoff condition," may include the following steps: If the number of skip connections in the intermediate first neural network that meets the iteration cutoff condition exceeds the second preset number, then the skip connection with the smallest weight in the intermediate first neural network is obtained; the skip connection with the smallest weight is replaced with the target neural network structure; the skip connection is a candidate neural network structure; the target neural network structure is the candidate neural network with the second largest weight in the intermediate first neural network.

[0099] In the process of determining the backbone network, when the iteration deadline is met, the intermediate first neural network that meets the deadline can be obtained. This intermediate first neural network generally includes at least two candidate neural network structures and their corresponding first weights. Then, the number of skip connections in the candidate neural network structures of this intermediate first neural network can be counted, and it can be determined whether the counted number of skip connections is greater than a second preset number. If it is greater, it means that there are too many skip connections in the intermediate first neural network, which will affect the recognition performance of the subsequently formed attribute recognition network. Therefore, the number of skip connections needs to be reduced. If the counted number of skip connections is not greater than the second preset number, the number of skip connections does not need to be reduced. The size of the second preset number can be set according to the actual situation, such as 2, 3, 4, etc.

[0100] When it is necessary to reduce the number of skip connections, the first weights of these candidate neural network structures in the intermediate first neural network can be sorted from largest to smallest, and the smallest first weight can be found from the sorting result. At the same time, the second largest (i.e., the second largest) first weight can also be found from the sorting result. Then, the candidate neural network structure / skip connection corresponding to the smallest first weight can be replaced with the candidate neural network structure / skip connection corresponding to the second largest first weight (denoted as the target neural network structure).

[0101] Since at least two candidate neural network structures in the aforementioned intermediate first neural network are generally not all skip connections, and the candidate neural network structure corresponding to the second largest first weight is generally not a skip connection, replacing the candidate neural network structure / skip connection corresponding to the smallest first weight with the candidate neural network structure corresponding to the second largest first weight (which is not a skip connection) can ensure the recognition performance of the subsequently formed attribute recognition network.

[0102] In this embodiment, when the number of skip connection network structures in the intermediate first neural network that meets the iteration cutoff condition is too large, some skip connections with smaller weights will be replaced with network structures with the second largest weights. This helps to maintain the simplicity of the backbone network, avoid overfitting of the final backbone network, and thus improve the recognition performance of the attribute recognition network determined by the backbone network.

[0103] After the backbone network structure is determined, one or more branch networks can be identified. Since branch networks have fewer parameters and are simpler to search than the backbone network, a gradient descent approach can be used instead of an incremental search strategy. The following example illustrates the process of determining branch networks using gradient descent.

[0104] In some embodiments, the step S206 above, "training all candidate neural network structures in a preset search space using gradient descent based on multiple training images to determine at least one branch network," may include the following steps: Step B1: Connect all candidate neural network structures in the preset search space to nodes as a network layer, and determine the initial branch network according to the second preset number of network layers.

[0105] In this step, the network layers can be constructed in the manner described in step A1 above. The number of layers in the branch network for identifying each attribute is determined according to the importance of each attribute of the sample object, and these are all denoted as the second preset layer number. Generally, the second preset layer number for important attributes is greater than the second preset layer number for less important attributes. That is to say, the second preset layer number for different attributes can be different, or they can be the same.

[0106] After setting the second preset number of layers and each network layer for each branch network, the network layers of the second preset number of layers of each branch network can be connected to obtain the connected branch networks, which are denoted as the initial branch networks.

[0107] Step B2: Based on the number of attributes of the sample objects to be identified, construct at least one initial branch network, and combine the backbone network and at least one initial branch network to determine the initial second neural network.

[0108] In this step, an initial branch network can be set up to match the number of attributes of the sample objects in the training images that need to be identified. The backbone network and at least one initial branch network can then be combined into a larger network, denoted as the initial second neural network.

[0109] For example, suppose we need to identify the gender and clothing style of a sample object, where gender is the most important attribute and clothing style is the least important attribute. We can set up two initial branch networks, Cell1 and Cell2, to identify the gender and clothing style of the sample object, respectively. Then, we can combine these two initial branch networks with the backbone network and other network layers to obtain the initial second neural network.

[0110] Step B3: Based on multiple training images, the initial second neural network is trained using gradient descent to obtain a converged second neural network; the second neural network includes at least one converged branch network.

[0111] In this step, after constructing the initial second neural network, a loss function for training the initial second neural network can be designed. Optionally, the loss function designed here can be a weighted loss function. Taking two branch networks as an example, the loss function shown in the following formula (7) can be designed: (7).

[0112] in, L all This represents the sum of all losses (denoted as the sum of losses); L 1. L 2 represents the loss of different branch networks, k 1. k 2 represents the weights of the loss in different branches of the network, and the sum of multiple loss weights is 1, for example... k 1+ k 2 = 1. Generally, the loss weight of important attributes is greater than the loss weight of less important attributes. For example, if gender is an important attribute and clothing style is a less important attribute, then the loss weight of gender is greater than that of less important attributes. k sex >Lower weighting of top style k upstyle .

[0113] After constructing the loss function of the initial second neural network, the structural parameters of the candidate neural network structures in each initial branch network of the initial second neural network can be initialized. Then, the initial second neural network can be trained using gradient descent as described in step A2 above to optimize the structural parameters / first weights of each candidate neural network structure until the initial second neural network converges. After that, the candidate neural network structure with the largest first weight in each initial branch network can be selected as its best-performing branch network structure.

[0114] In conjunction with the above description of the loss function, the training process can optionally include: Multiple training images are input into the initial second neural network for attribute recognition processing to determine at least two predicted attributes corresponding to the sample objects in each training image. For each training image, the loss between each predicted attribute and the corresponding labeled attribute of the sample object in the training image is calculated, and the loss is weighted and summed to determine the loss sum value corresponding to the training image. Based on the loss sum value of each training image, the initial second neural network is trained using gradient descent to obtain a converged second neural network.

[0115] The loss between each predicted attribute and its corresponding labeled attribute can be calculated using the cross-entropy loss function in formula (6) above. Of course, other loss calculation methods can also be used. In short, the loss between each predicted attribute and its corresponding labeled attribute can be calculated. Then, the losses can be weighted and summed according to formula (7) above to obtain the sum of losses. The parameters are backpropagated through the sum of losses, and the initial second neural network is trained by combining gradient descent to finally obtain a converged second neural network.

[0116] In this embodiment, gradient descent is used to train the search branch networks, which reduces search time, improves network efficiency, and better utilizes resources. Furthermore, by employing a weighted loss function during branch network search training, the final attribute recognition network is more biased towards more important object attributes, ensuring that important attributes are prioritized within limited performance constraints.

[0117] The following explains how the final attribute recognition network is obtained based on the backbone network and branch networks determined above.

[0118] In some embodiments, the step of "determining the attribute identification network based on the backbone network and at least one branch network" in S208 above may include the following steps: The backbone network and at least one branch network are combined to determine the initial attribute recognition network; the training set and validation set of the sample objects are combined into a new training set, and the initial attribute recognition network is trained using the new training set to determine the attribute recognition network.

[0119] Among them, the backbone network and each branch network determined above can be arranged according to the above... Figure 5 The network obtained by combining the methods described above is denoted as the initial attribute recognition network. See also... Figure 6 The schematic diagram shown illustrates the overall process of defining the attribute recognition network. The training and validation sets defined in step S202 can be merged to form a new training set. The initial attribute recognition network can then be retrained using this new training set. After retraining, the performance of the trained attribute recognition network can be tested using the aforementioned test set.

[0120] The above retraining method can be adopted in the following two ways: The first training method is to label multiple attribute tags (i.e., label attributes) for each training image in the new training set according to the number of attributes to be recognized in a certain order. When retraining the initial attribute recognition network, the network parameters of the backbone network in the initial attribute recognition network are not fixed. Instead, the network parameters in the initial attribute recognition network are retrained by combining the weighted loss function in the branch network mentioned above, and finally a well-trained attribute recognition network is obtained.

[0121] The second training method involves assigning a unique attribute label to each training image in the new training set. Training images for different attributes can be reused, and the more important and generalizable attributes are selected as training data for the backbone network parameters. After training, the backbone network parameters are frozen, and each branch network is trained sequentially using the cross-entropy loss function to finally obtain the trained attribute recognition network.

[0122] Generally, under the same test set, the first training method will outperform the second in terms of test metrics. However, the second method is more flexible. For example, when reducing or adding branches and retraining, the first method requires relabeling, while the second method does not. The specific training method used for retraining can be chosen based on the actual situation.

[0123] In this embodiment, the initial attribute recognition network, which consists of the backbone network and the branch networks of the search, is retrained using a new training set, which can improve the recognition performance of the finally determined attribute recognition network.

[0124] The object attribute recognition device provided by the present invention will be described below. The object attribute recognition device described below and the object attribute recognition method described above can be referred to each other. The detailed implementation process of the device and the method will not be repeated here.

[0125] Figure 7 This is a schematic diagram of the object attribute recognition device provided in an embodiment of the present invention. See also: Figure 7 As shown, the device may include: The acquisition module 310 is used to acquire an image to be recognized; the image to be recognized includes an object to be recognized. The attribute recognition module 320 is used to input the image to be recognized into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be recognized; The aforementioned attribute recognition network is determined by searching for the candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

[0126] In some embodiments, the above-described apparatus may further include a determining module, which may include: The acquisition unit is used to acquire multiple training images; each training image includes sample objects with labeled attributes, and the multiple training images constitute the training set, validation set, and test set of the sample objects; The backbone network search unit is used to train all candidate neural network structures in the preset search space based on multiple training images using gradient descent and progressive search strategies to determine the backbone network. The branch network search unit is used to train all candidate neural network structures in a preset search space using gradient descent based on multiple training images, and to determine at least one branch network; each branch network is used to identify one attribute of the sample object. A determining unit is used to determine the attribute identification network based on the backbone network and at least one branch network.

[0127] Optionally, the aforementioned backbone network search unit is specifically used for

[0128] All candidate neural network structures in the preset search space are connected to nodes as a network layer, and the initial first neural network is determined according to the network layer with the first preset number of layers. The training operation includes: training an initial first neural network using gradient descent based on multiple training images to obtain a converged first neural network; the first neural network includes trained candidate neural network structures and first weights corresponding to each candidate neural network structure. Perform a progressive search operation, which includes: based on the first weight of each candidate neural network structure, using a progressive search strategy to delete a first preset number of candidate neural network structures in the first neural network, and increasing the number of layers in the first neural network to obtain an intermediate first neural network; The intermediate first neural network is used as the new initial first neural network. The above training operation and progressive search operation are performed iteratively until the preset iteration cutoff condition is met. The backbone network is obtained based on the intermediate first neural network when the iteration cutoff condition is met.

[0129] Optionally, the aforementioned backbone network search unit is specifically used for

[0130] If the number of skip connections in the intermediate first neural network that meets the iteration cutoff condition exceeds the second preset number, then the skip connection with the smallest weight in the intermediate first neural network is obtained; the skip connection is a candidate neural network structure; the skip connection with the smallest weight is replaced with the target neural network structure; the target neural network structure is the candidate neural network structure with the second largest weight in the intermediate first neural network.

[0131] Optionally, the aforementioned branch network search unit is specifically used for

[0132] All candidate neural network structures in the preset search space are connected to nodes as a network layer, and the initial branch network is determined according to the network layer with the second preset number of layers. Based on the number of attributes of the sample objects to be identified, at least one initial branch network is constructed, and the backbone network and at least one initial branch network are combined to determine the initial second neural network; Based on multiple training images, an initial second neural network is trained using gradient descent to obtain a converged second neural network; the second neural network includes at least one converged branch network.

[0133] Optionally, the aforementioned branch network search unit is specifically used for

[0134] Multiple training images are input into the initial second neural network for attribute recognition processing to determine at least two predicted attributes corresponding to the sample objects in each training image. For each training image, the loss between each predicted attribute and the corresponding labeled attribute of the sample object in the training image is calculated, and the loss is weighted and summed to determine the loss sum value corresponding to the training image. Based on the loss sum value of each training image, the initial second neural network is trained using gradient descent to obtain a converged second neural network.

[0135] In some embodiments, the determining unit is specifically used for

[0136] The backbone network and at least one branch network are combined to determine the initial attribute recognition network; The training and validation sets of the sample objects are combined into a new training set, and the initial attribute recognition network is trained using the new training set to determine the attribute recognition network.

[0137] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute an object attribute recognition method. This method includes: acquiring an image to be recognized; the image to be recognized includes an object to be recognized; inputting the image to be recognized into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be recognized; wherein the attribute recognition network is determined by searching for candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

[0138] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program that can be stored on a computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the object attribute recognition method provided by the above methods. The method includes: acquiring an image to be recognized; the image to be recognized including an object to be recognized; inputting the image to be recognized into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be recognized; wherein the attribute recognition network is determined by searching for candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

[0140] In another aspect, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the object attribute recognition method provided by the above methods. The method includes: acquiring an image to be recognized; the image to be recognized including an object to be recognized; inputting the image to be recognized into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be recognized; wherein the attribute recognition network is determined by searching for candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

[0141] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0142] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying object attributes, characterized in that, include: Acquire the image to be recognized; The image to be identified includes an object to be identified; The image to be identified is input into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be identified; The attribute recognition network is determined by searching for the candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

2. The object attribute recognition method according to claim 1, characterized in that, The method for determining the attribute recognition network includes: Multiple training images are acquired; each training image includes a sample object with labeled attributes, and the multiple training images constitute the training set, validation set, and test set of the sample object; Based on the multiple training images, gradient descent and progressive search strategies are used to train all candidate neural network structures in the preset search space to determine the backbone network. Based on the multiple training images, gradient descent is used to train all candidate neural network structures in the preset search space to determine at least one branch network; each branch network is used to identify one attribute of the sample object. The attribute identification network is determined based on the backbone network and the at least one branch network.

3. The object attribute recognition method according to claim 2, characterized in that, The step of training all candidate neural network structures in the preset search space using gradient descent and a progressive search strategy based on the multiple training images to determine the backbone network includes: All candidate neural network structures in the preset search space are connected to nodes as a network layer, and an initial first neural network is determined according to the network layer of the first preset number of layers. Perform a training operation, the training operation including: training the initial first neural network using gradient descent based on the plurality of training images to obtain a converged first neural network; the first neural network includes the trained candidate neural network structure and a first weight corresponding to each candidate neural network structure; Perform a progressive search operation, the progressive search operation including: according to the first weight of each candidate neural network structure, using the progressive search strategy to delete a first preset number of candidate neural network structures in the first neural network, and increasing the number of layers of the first neural network to obtain an intermediate first neural network; The intermediate first neural network is used as a new initial first neural network. The training operation and the progressive search operation are executed iteratively until the preset iteration cutoff condition is met. The backbone network is obtained based on the intermediate first neural network when the iteration cutoff condition is met.

4. The object attribute recognition method according to claim 3, characterized in that, The step of obtaining the backbone network based on the intermediate first neural network when the iteration cutoff condition is met includes: If the number of skip connections in the intermediate first neural network that meets the iteration cutoff condition exceeds the second preset number, then the skip connection with the smallest weight in the intermediate first neural network is obtained; the skip connection is a candidate neural network structure. Replace the first skip connection with the smallest weight with the target neural network structure; the target neural network structure is the candidate neural network structure with the second largest weight in the intermediate first neural network.

5. The object attribute recognition method according to claim 2, characterized in that, The step of training all candidate neural network structures in the preset search space using gradient descent based on the multiple training images to determine at least one branch network includes: All candidate neural network structures in the preset search space are connected to nodes as a network layer, and the initial branch network is determined according to the network layer of the second preset number of layers. Based on the number of attributes of the sample objects to be identified, at least one initial branch network is constructed, and the backbone network and the at least one initial branch network are combined to determine an initial second neural network; Based on the multiple training images, the initial second neural network is trained using gradient descent to obtain a converged second neural network; the second neural network includes at least one converged branch network.

6. The object attribute recognition method according to claim 5, characterized in that, The step of training the initial second neural network using gradient descent based on the plurality of training images to obtain a converged second neural network includes: The plurality of training images are respectively input into the initial second neural network for attribute recognition processing to determine at least two predicted attributes corresponding to the sample object in each training image; For each training image, calculate the loss between each predicted attribute and the corresponding labeled attribute of the sample object in the training image, and perform a weighted summation of each loss to determine the loss sum value corresponding to the training image; The initial second neural network is trained using gradient descent based on the loss and value of each training image to obtain a converged second neural network.

7. The object attribute identification method according to any one of claims 2 to 6, characterized in that, Determining the attribute identification network based on the backbone network and the at least one branch network includes: The backbone network and the at least one branch network are combined to determine the initial attribute recognition network; The training set and validation set of the sample objects are combined into a new training set, and the initial attribute recognition network is trained using the new training set to determine the attribute recognition network.

8. An object attribute recognition device, characterized in that, include: The acquisition module is used to acquire the image to be recognized; The image to be identified includes an object to be identified; An attribute recognition module is used to input the image to be recognized into a preset attribute recognition network for attribute recognition processing to determine at least one attribute of the object to be recognized; The attribute recognition network is determined by searching for the candidate neural network structures in a preset search space using gradient descent and a progressive search strategy, and the preset search space includes at least one candidate neural network structure.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the object attribute recognition method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the object attribute identification method as described in any one of claims 1 to 7.

Images