Target recognition method and device based on feature saliency and computer device

Abstract

The application relates to a target recognition method and device based on feature saliency and a computer device. An extreme learning sparse self-encoding network is constructed and trained, wherein the number of hidden layer nodes in the extreme learning sparse self-encoding network is less than the number of input layer nodes and the number of output layer nodes; a saliency component extraction network is constructed according to the input layer and the hidden layer of the trained extreme learning sparse self-encoding network, convolution features are input into the saliency component extraction network to obtain saliency components of the convolution features, and saliency features are obtained by fusing the saliency components and the convolution features; and the saliency features are input into a pre-trained recognition model to obtain a target recognition result. The application can balance the accuracy and the calculation speed of the recognition result.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a target recognition method, apparatus and computer device based on feature saliency. Background Technology

[0002] Accurate target recognition is crucial for mobile robots to perform tasks such as perception, decision-making, planning, and control, and it represents a fundamental research area in mobile robot environmental perception technology. Features are the prerequisite for effective and robust target recognition. However, current feature engineering and feature learning methods, especially the mainstream deep neural network-based feature learning methods, still suffer from insufficient discriminative generalization ability; that is, recognition accuracy is unstable when faced with data outside the training data. Most current mainstream methods rely on prior knowledge or engineering experience to improve existing mainstream deep neural network structures by designing and adding functional network modules relevant to specific tasks, hoping to obtain more representative features and thus improve target recognition performance. However, this approach often comes at the cost of a significant increase in network complexity, leading to a sharp increase in data requirements, training costs, and model computation time, resulting in a decrease in real-time target recognition performance. Summary of the Invention

[0003] Therefore, it is necessary to provide a target recognition method, apparatus, and computer device based on feature saliency to address the aforementioned technical problems, so as to improve both the accuracy and speed of target recognition.

[0004] A target recognition method based on feature saliency, the method comprising:

[0005] Obtain the convolutional features of the image to be identified;

[0006] Construct and train an extreme learning sparse autoencoder network; the number of hidden layer nodes in the extreme learning sparse autoencoder network is less than the number of input layer nodes and the number of output layer nodes;

[0007] A salient component extraction network is constructed based on the input layer and hidden layer of the trained Extreme Learning Sparse Autoencoder Network. The convolutional features are input into the salient component extraction network to obtain the salient components of the convolutional features. The salient components and the convolutional features are then fused to obtain the salient features.

[0008] The salient features are input into a pre-trained recognition model to obtain the target recognition result.

[0009] In one embodiment, the process of training an extreme learning sparse autoencoder network includes:

[0010] Obtain the training set of convolutional features;

[0011] The convolutional feature training set is input into the Extreme Learning Sparse Autoencoder Network, and the first objective function of the Extreme Learning Sparse Autoencoder Network is constructed using the activation output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set.

[0012] Solving the first objective function yields the trained Limit Learning Sparse Autoencoder Network.

[0013] In one embodiment, a first objective function of the extreme learning sparse autoencoder network is constructed using the activation output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set, including:

[0014] The first objective function of the Extreme Learning Sparse Autoencoder network, constructed using the activation output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set, is as follows:

[0015]

[0016] Where, β * Let A represent the optimal output weight matrix when the objective function is optimal. F X represents the activation-output matrix of the hidden layer. F Let β represent the training set of convolutional features, and let β represent the output weight matrix. This indicates the calculation of the l1 norm.

[0017] In one embodiment, the convolutional features are input into a salient component extraction network to obtain salient components of the convolutional features, and the salient components and the convolutional features are fused to obtain salient features, including:

[0018] The convolutional features are input into a saliency component extraction network to obtain the saliency components of the convolutional features. The saliency components of the convolutional features are then obtained based on the optimal output weight matrix.

[0019]

[0020] in, x represents the saliency component of the convolutional feature. CNN Represents convolutional features;

[0021] The salient features are obtained by fusing the salient components and the convolutional features:

[0022]

[0023] Where, x Fc Indicates a significant feature.

[0024] In one embodiment, before obtaining the convolutional features of the image to be identified, the following steps are included:

[0025] Obtain the image training sample set;

[0026] The image training sample set is input into a pre-constructed convolutional neural network;

[0027] The cost function for constructing the convolutional neural network is:

[0028]

[0029] in, Let f(x) be the label prediction loss term for the training samples. i ) represents the convolutional feature extraction network relative to the input i-th training sample x. i The corresponding network output, y i Let represent the category label corresponding to the i-th training sample, and N represent the number of samples contained in the image training sample set. λ represents the network weight regularization term, λ is the weighting parameter of the network weight regularization term, and k is the number of layers in the neural network.

[0030] The cost function is optimized to obtain a trained convolutional neural network. The fully connected layers in the trained convolutional neural network are removed to obtain the convolutional feature extraction backbone network.

[0031] In one embodiment, the training process of the recognition model includes:

[0032] Obtain a training sample set with significant features;

[0033] The training sample set of the salient features is input into the recognition model to be trained to construct the second objective function of the recognition model to be trained:

[0034]

[0035] Where α represents the output weight matrix of the output layer of the recognition model to be trained, α * The optimal output weight matrix, where C is the regularization coefficient, and A... Fc Let L be the hidden activation output matrix of the recognition model to be trained, and L be the training sample label matrix.

[0036] Solving the second objective function yields the optimal output weight matrix of the output layer of the trained recognition model:

[0037]

[0038] In one embodiment, solving the first objective function to obtain the trained extreme learning sparse autoencoder network includes:

[0039] The first objective function is solved using a fast iterative threshold shrinkage algorithm to obtain a trained extreme learning sparse autoencoder network.

[0040] In one embodiment, the recognition model is an extreme learning machine.

[0041] A target recognition device based on feature saliency, the device comprising:

[0042] The acquisition module is used to acquire the convolutional features of the image to be recognized;

[0043] The training module is used to construct and train an extreme learning sparse autoencoder network; the number of hidden layer nodes in the extreme learning sparse autoencoder network is less than the number of input layer nodes and the number of output layer nodes.

[0044] The fusion module is used to construct a salient component extraction network based on the input layer and hidden layer of the trained Extreme Learning Sparse Autoencoder Network, input the convolutional features into the salient component extraction network to obtain the salient components of the convolutional features, and fuse the salient components and the convolutional features to obtain salient features.

[0045] The recognition module is used to input the salient features into a pre-trained recognition model to obtain the target recognition result.

[0046] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0047] Obtain the convolutional features of the image to be identified;

[0048] Construct and train an extreme learning sparse autoencoder network; the number of hidden layer nodes in the extreme learning sparse autoencoder network is less than the number of input layer nodes and the number of output layer nodes;

[0049] A salient component extraction network is constructed based on the input layer and hidden layer of the trained Extreme Learning Sparse Autoencoder Network. The convolutional features are input into the salient component extraction network to obtain the salient components of the convolutional features. The salient components and the convolutional features are then fused to obtain the salient features.

[0050] The salient features are input into a pre-trained recognition model to obtain the target recognition result.

[0051] The aforementioned target recognition method, apparatus, and computer equipment based on feature saliency construct and train an Extreme Learning Sparse Autoencoder (ELA) network, wherein the number of hidden layer nodes in the ELA is less than the number of input layer nodes and output layer nodes. A saliency component extraction network is constructed based on the input and hidden layers of the trained ELA. Convolutional features are input into the saliency component extraction network to obtain the saliency components of the convolutional features. The saliency components and convolutional features are fused to obtain salient features. The salient features are then input into a pre-trained recognition model to obtain the target recognition result. The ELA of this invention performs unsupervised extraction of saliency components from convolutional features extracted through supervised learning. By designing the number of hidden layer nodes, the saliency component extraction process from convolutional features is modeled as a sparse, low-dimensional transformation reconstructing the original convolutional features. The addition of sparsity constraints greatly improves the generalization ability of salient features. Furthermore, the ELA borrows the optimization principles of Extreme Learning Machines, significantly saving computational resources and improving computational speed. Overall, this invention can balance the accuracy of the recognition result with computational speed. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a target recognition method based on feature saliency in one embodiment;

[0053] Figure 2 This is a schematic diagram of the architecture of an extreme learning sparse autoencoder network in one embodiment;

[0054] Figure 3 This is a method for constructing a salient component extraction network and a network architecture diagram in one embodiment;

[0055] Figure 4 This is a schematic diagram of the overall framework of a target recognition method based on salient features in another embodiment;

[0056] Figure 5 This is a structural block diagram of a target recognition device based on feature saliency in one embodiment;

[0057] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1 As shown, a target recognition method based on feature saliency is provided, including the following steps:

[0060] Step 102: Obtain the convolutional features of the image to be identified.

[0061] Convolutional features are obtained by inputting the image to be recognized into a pre-trained convolutional neural network. It is worth noting that the convolutional features can be any one or more layers of features from the multi-level image features extracted by the convolutional neural network, and can be selected according to the actual situation.

[0062] This approach does not require the original image as input. On the one hand, it reduces the amount of computation and avoids processing too much redundant information and the influence of interfering information. On the other hand, it facilitates the fusion of (single-layer or multi-layer) convolutional feature information and typical image feature information (such as Hog) and the extraction of saliency information. This compensates for the loss of beneficial feature information in the original CNN network during the feature extraction process layer by layer, and further improves the performance of subsequent classification and recognition.

[0063] Step 104: Construct and train the Extreme Learning Sparse Autoencoder Network.

[0064] Extreme Learning Sparse Autoencoder Network (SELM-AE) consists of an input layer, hidden layers, and an output layer, such as... Figure 2 The diagram shows the architecture of the Extreme Learning Sparse Autoencoder Network, which includes d input layer nodes. There are d hidden layer nodes and d output layer nodes. The number of nodes in the hidden layer is less than the number of nodes in the input layer and the number of nodes in the output layer. The hidden layers are feature encoding layers. The input and hidden layers constitute the Extreme Learning Machine (ELM) sparse encoder, and the hidden and output layers constitute the Extreme Learning Machine (ELM) sparse decoder. ELM stands for Extreme Learning Machine, S represents sparse, AE represents autoencoder, and β represents the output weight matrix of the output layer. The Extreme Learning Sparse Autoencoder network borrows the optimization principles of the Extreme Learning Machine, significantly saving computational resources and improving computational speed.

[0065] Step 106: Construct a salient component extraction network based on the input layer and hidden layer of the trained Extreme Learning Sparse Autoencoder Network. Input the convolutional features into the salient component extraction network to obtain the salient components of the convolutional features. Fuse the salient components and convolutional features to obtain salient features.

[0066] Features learned from deep neural networks still have room for improvement in saliency. There are still feature components that, compared to other feature components, are more beneficial for enhancing the discriminative power between samples and improving the final classification accuracy—these are called saliency components. This invention posits that feature saliency refers to the degree to which a sample feature best reflects the essential attributes of the sample. The saliency components contained in a feature can, to a certain extent, be regarded as components or combinations of components in the feature vector that have a significant impact on the discriminative power between samples and the final classification accuracy. Such components or their combinations should possess stronger representational and generalization abilities. Naturally, if such saliency components can be learned and extracted through some means and applied to target classification tasks, the corresponding recognition performance should be further improved.

[0067] like Figure 3 As shown, a method for constructing a salient component extraction network and a network architecture diagram are provided, where x represents the input convolutional features. The output represents the salient components. As can be seen, the input and hidden layers of the trained Extreme Learning Sparse Autoencoder network serve as the input and output layers of the salient component extraction network, respectively. In this network, the number of nodes in the output layer is less than the number of nodes in the input layer. This is equivalent to modeling the salient component extraction process from convolutional features as a sparse, low-dimensional transformation that reconstructs the original convolutional features. Unsupervised learning, represented by autoencoders, requires no labels and aims to fully explore the inherent patterns and inspirations within the data itself. Furthermore, by incorporating sparsity constraints, the generalization ability of salient features is significantly improved.

[0068] Step 108: Input the salient features into the pre-trained recognition model to obtain the target recognition result.

[0069] The aforementioned target recognition method based on feature saliency constructs and trains an Extreme Learning Sparse Autoencoder (ELA) network, where the number of hidden layer nodes is less than the number of input layer nodes and output layer nodes. A saliency component extraction network is then built based on the input and hidden layers of the trained EIA network. Convolutional features are input into this network to obtain saliency components, and the saliency components and convolutional features are fused to obtain salient features. These salient features are then input into a pre-trained recognition model to obtain the target recognition result. This method's EIA network performs unsupervised extraction of saliency components from convolutional features extracted through supervised learning. By designing the number of hidden layer nodes, the saliency component extraction process is modeled as a sparse, low-dimensional transformation of the original convolutional features. The addition of sparsity constraints significantly improves the generalization ability of salient features. Furthermore, the EIA network borrows optimization principles from Extreme Learning Machines, greatly saving computational resources and improving computational speed. This method combines the advantages of both supervised and unsupervised learning. By extracting the salient components of convolutional features and further fusing them with the convolutional features before inputting them into the recognition model, it can achieve fast and accurate target recognition.

[0070] In one embodiment, before obtaining the convolutional features of the image to be identified, the following steps are included:

[0071] Obtain an image training sample set, input the image training sample set into a pre-built convolutional neural network, and construct the cost function of the convolutional neural network:

[0072]

[0073] in, Let f(x) be the label prediction loss term for the training samples. i ) represents the convolutional feature extraction network relative to the input i-th training sample x. i The corresponding network output, y i Let represent the category label corresponding to the i-th training sample, and N represent the number of samples contained in the image training sample set. λ represents the network weight regularization term, which is the weight parameter of the network weight regularization term and can affect the sparsity of the network weights. k is the number of layers in the neural network.

[0074] The cost function is optimized to obtain a trained convolutional neural network. The fully connected layers in the trained convolutional neural network are then removed to obtain the convolutional feature extraction backbone network.

[0075] The process described above in this embodiment is the training and construction process of the convolutional feature extraction backbone network used to extract convolutional features. Its function is to learn and extract the corresponding deep convolutional features from the input original image pixels. This process is mainly achieved by training the corresponding deep convolutional network model on the training dataset related to the target recognition task. The fully connected classification layer of the deep convolutional network obtained by training will be removed, and only the backbone network part related to feature extraction will be retained. Currently representative deep convolutional network models can be used, such as the VGG-M model.

[0076] In one embodiment, the process of training an extreme learning sparse autoencoder network includes:

[0077] Obtain the convolutional feature training set, input the convolutional feature training set into the Extreme Learning Sparse Autoencoder Network, and construct the first objective function of the Extreme Learning Sparse Autoencoder Network using the activation-output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set:

[0078]

[0079] Where, β * Let A represent the optimal output weight matrix when the objective function is optimal. F X represents the activation-output matrix of the hidden layer. F Let β represent the training set of convolutional features, and let β represent the output weight matrix. This indicates the calculation of the l1 norm.

[0080] The trained extreme learning sparse autoencoder network is obtained by solving the first objective function. The first objective function can be solved using a fast iterative threshold shrinkage algorithm.

[0081] This process is mainly achieved by extracting deep convolutional features from the training dataset related to the target recognition task, and then using the obtained deep convolutional feature set as input to train the SELM-AE network.

[0082] In one embodiment, convolutional features are input into a salient component extraction network to obtain salient components of the convolutional features, and the salient components and convolutional features are fused to obtain salient features, including:

[0083] The convolutional features are input into a saliency component extraction network to obtain the saliency components of the convolutional features. The saliency components of the convolutional features are then obtained based on the optimal output weight matrix.

[0084]

[0085] in, x represents the saliency component of the convolutional feature. CNN Represents convolutional features;

[0086] The salient features are obtained by fusing salient components and convolutional features:

[0087]

[0088] Where, x Fc Indicates a significant feature.

[0089] In one embodiment, the recognition model is an extreme learning machine. The extreme learning machine transforms the training of a single hidden layer feedforward neural network into solving the least squares solution of a specific linear system. It replaces the traditional iterative gradient descent with analytical matrix operations, thus enabling the application of the extreme learning machine principle to train neural networks. This has outstanding advantages such as fast computation to meet real-time learning needs, fewer learning parameters that need to be manually set, high accuracy, and good generalization ability.

[0090] The training process of the recognition model includes: obtaining a training sample set of salient features, inputting the training sample set of salient features into the recognition model to be trained, and constructing the second objective function of the recognition model to be trained.

[0091]

[0092] Where α represents the output weight matrix of the output layer of the recognition model to be trained, α * The optimal output weight matrix, where C is the regularization coefficient, and A... Fc Let L be the hidden activation output matrix of the recognition model to be trained, and L be the training sample label matrix.

[0093] Solving the second objective function yields the optimal output weight matrix of the output layer of the trained recognition model:

[0094]

[0095] During the classifier learning and training process, the parameters to be set are mainly the number of hidden layer nodes of the ELM classifier and the corresponding activation function type. Commonly used ones include the Sigmoid function, Gaussian function, and hyperbolic tangent function.

[0096] like Figure 4 The diagram shows the overall framework of a target recognition method based on feature saliency.

[0097] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0098] In one embodiment, such as Figure 5 As shown, a target recognition device based on feature saliency is provided, comprising: an acquisition module, a training module, a fusion module, and a recognition module, wherein:

[0099] The acquisition module is used to acquire the convolutional features of the image to be recognized;

[0100] The training module is used to build and train the Extreme Learning Sparse Autoencoder Network; the number of hidden layer nodes in the Extreme Learning Sparse Autoencoder Network is less than the number of input layer nodes and the number of output layer nodes.

[0101] The fusion module is used to construct a salient component extraction network based on the input layer and hidden layer of the trained limit learning sparse autoencoder network. The convolutional features are input into the salient component extraction network to obtain the salient components of the convolutional features. The salient components and convolutional features are then fused to obtain the salient features.

[0102] The recognition module is used to input salient features into a pre-trained recognition model to obtain target recognition results.

[0103] In one embodiment, the training module is further configured to obtain a convolutional feature training set, input the convolutional feature training set into the Extreme Learning Sparse Autoencoder Network, and construct the first objective function of the Extreme Learning Sparse Autoencoder Network using the activation output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set:

[0104]

[0105] Where, β * Let A represent the optimal output weight matrix when the objective function is optimal. F X represents the activation-output matrix of the hidden layer. F Let β represent the training set of convolutional features, and let β represent the output weight matrix. This indicates the calculation of the l1 norm.

[0106] The first objective function is solved using a fast iterative threshold shrinkage algorithm to obtain a trained extreme learning sparse autoencoder network.

[0107] In one embodiment, the fusion module is further configured to input the convolutional features into a salient component extraction network to obtain the salient components of the convolutional features, and to obtain the salient components of the convolutional features based on the optimal output weight matrix.

[0108]

[0109] in, x represents the saliency component of the convolutional feature. CNN Represents convolutional features;

[0110] The salient features are obtained by fusing salient components and convolutional features:

[0111]

[0112] Where, x Fc Indicates a significant feature.

[0113] For specific limitations regarding the feature-saliency-based target recognition device, please refer to the limitations of the feature-saliency-based target recognition method above, which will not be repeated here. Each module in the aforementioned feature-saliency-based target recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0114] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores image data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a target recognition method based on feature saliency.

[0115] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0116] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0117] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0118] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0119] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0120] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A target recognition method based on feature saliency, characterized in that, The method includes: Obtain the convolutional features of the image to be identified; Construct and train an extreme learning sparse autoencoder network; the number of hidden layer nodes in the extreme learning sparse autoencoder network is less than the number of input layer nodes and the number of output layer nodes; A salient component extraction network is constructed based on the input layer and hidden layer of the trained Extreme Learning Sparse Autoencoder Network. The convolutional features are input into the salient component extraction network to obtain the salient components of the convolutional features. The salient components and the convolutional features are then fused to obtain the salient features. The salient features are input into a pre-trained recognition model to obtain the target recognition result.

2. The method according to claim 1, characterized in that, The process of training a Extreme Learning Sparse Autoencoder Network includes: Obtain the training set of convolutional features; The convolutional feature training set is input into the Extreme Learning Sparse Autoencoder Network, and the first objective function of the Extreme Learning Sparse Autoencoder Network is constructed using the activation output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set. Solving the first objective function yields the trained Limit Learning Sparse Autoencoder Network.

3. The method according to claim 2, characterized in that, The first objective function of the extreme learning sparse autoencoder network is constructed using the activation output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set, including: The first objective function of the Extreme Learning Sparse Autoencoder network, constructed using the activation output matrix of the hidden layer, the output weight matrix of the output layer, and the input convolutional feature training set, is as follows: in, This represents the optimal output weight matrix when the objective function is optimal. This represents the activation-output matrix of the hidden layer. This represents the training set of convolutional features. This represents the output weight matrix. Indicates calculation Norm.

4. The method according to claim 3, characterized in that, The convolutional features are input into a salient component extraction network to obtain the salient components of the convolutional features. The salient components and the convolutional features are then fused to obtain salient features, including: The convolutional features are input into a saliency component extraction network to obtain the saliency components of the convolutional features. The saliency components of the convolutional features are then obtained based on the optimal output weight matrix. in, The salient components representing convolutional features. Represents convolutional features; The salient features are obtained by fusing the salient components and the convolutional features: in, Indicates a significant feature.

5. The method according to claim 1, characterized in that, Before obtaining the convolutional features of the image to be recognized, the following steps are included: Obtain the image training sample set; The image training sample set is input into a pre-constructed convolutional neural network; The cost function for constructing the convolutional neural network is: in, The loss term is the label prediction term for the training samples. This indicates that the convolutional feature extraction network is relative to the input... training samples The corresponding network output, Indicates the first The category label corresponding to each training sample This indicates the number of samples contained in the image training sample set. This represents the network weight regularization term. This represents the weighting parameter for the network weight regularization term. The number of layers in the neural network; The cost function is optimized to obtain a trained convolutional neural network. The fully connected layers in the trained convolutional neural network are removed to obtain the convolutional feature extraction backbone network.

6. The method according to claim 1, characterized in that, The training process of the recognition model includes: Obtain a training sample set with significant features; The training sample set of the salient features is input into the recognition model to be trained to construct the second objective function of the recognition model to be trained: in, This represents the output weight matrix of the output layer of the recognition model to be trained. Optimal output weight matrix, The regularization coefficient is . Let be the hidden activation output matrix of the recognition model to be trained. The training sample label matrix; Solving the second objective function yields the optimal output weight matrix of the output layer of the trained recognition model: 。 7. The method according to claim 1, characterized in that, Solving the first objective function yields the trained extreme learning sparse autoencoder network, including: The first objective function is solved using a fast iterative threshold shrinkage algorithm to obtain a trained extreme learning sparse autoencoder network.

8. The method according to claim 1, characterized in that, The recognition model is an extreme learning machine.

9. A target recognition device based on feature saliency, characterized in that, The device includes: The acquisition module is used to acquire the convolutional features of the image to be recognized; The training module is used to construct and train an extreme learning sparse autoencoder network; the number of hidden layer nodes in the extreme learning sparse autoencoder network is less than the number of input layer nodes and the number of output layer nodes. The fusion module is used to construct a salient component extraction network based on the input layer and hidden layer of the trained Extreme Learning Sparse Autoencoder Network, input the convolutional features into the salient component extraction network to obtain the salient components of the convolutional features, and fuse the salient components and the convolutional features to obtain salient features. The recognition module is used to input the salient features into a pre-trained recognition model to obtain the target recognition result.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images