Interpretable target recognition methods, devices, and equipment based on deep sparse coding models

Abstract

This application relates to an interpretable target recognition method, apparatus, and device based on a deep sparse coding model. The method first performs sparse decomposition on the SAR image to be identified using a convolutional sparse coding model based on a deep unfolded network to extract sparse features representing low-dimensional target features. Then, a feature learning model is used to obtain discriminative feature representations from the sparse features and extract latent category semantic features from them. Finally, a classification model identifies the target in the SAR image. This method achieves accurate target classification and recognition in SAR images while enhancing the interpretability of the model during image processing.

Description

Technical Field

[0001] This application relates to the field of SAR image target recognition technology, and in particular to an interpretable target recognition method, apparatus and device based on a deep sparse coding model. Background Technology

[0002] In recent years, deep learning technology has flourished and has been widely applied in the field of SAR image interpretation. However, deep learning technology is data-driven, achieving specific tasks by fitting implicit objective functions in a huge parameter space. Its working mechanism and decision-making logic are difficult to understand, making it seem like a black box.

[0003] Current deep learning model interpretability can be categorized into two types: post-hoc interpretability and pre-hoc interpretability. Post-hoc interpretability uses features visualization and case studies to explain existing models. For example, for class activation maps, it utilizes class-related parameter information or gradient information in network layers to visualize the response of the original image in the network model, effectively explaining what the model "focuses on." In induction motor fault signal classification, the Layer-wise Relevance Propagation (LRP) algorithm generates pixel-level explanations, better demonstrating the impact of key slices in the signal on the diagnostic results. Since prediction and interpretation are two independent processes, the advantage of post-hoc interpretability is that it does not affect the model's predictive performance. However, post-hoc interpretability often deviates from the true model, and this risk is difficult to avoid. Pre-hoc interpretability, on the other hand, often proactively modifies the network structure or training process based on physical or mathematical models to achieve better interpretability and model transparency. Some existing technologies use the prediction results of physical models as constraints for model learning, proposing a physical learning framework for water temperature inversion. Alternatively, one could embed successive wavelet convolutional layers within a CNN framework to extract interpretable features. Methods based on pre-interpretability imply introducing more regularization constraints or interpretable theoretical models into deep networks to obtain more interpretable features.

[0004] In the field of SAR target recognition, the interpretability of deep learning models has attracted increasing attention. Introducing SAR-related prior knowledge into deep networks to assist decision-making not only improves the model's generalization performance but also enhances its interpretability. Embedding electromagnetic scattering information provided by attribute scattering center models into deep learning models effectively improves the robustness of SAR target recognition models. Based on this, various related deep learning models have been proposed for SAR target recognition in existing technologies. However, these models lack deep integration of expert knowledge / theoretical models into the deep network learning process, and thus still cannot effectively improve the model's interpretability. Summary of the Invention

[0005] Therefore, it is necessary to provide an interpretable target recognition method, apparatus, and device based on a deep sparse coding model that can improve interpretability while possessing accurate recognition capabilities, in order to address the aforementioned technical problems.

[0006] An interpretable target recognition method based on a deep sparse coding model, the method comprising:

[0007] Acquire SAR images for target identification;

[0008] The SAR image is input into a convolutional sparse coding model based on a deep unfolding network. The model performs sparse decomposition on the SAR image and extracts sparse features representing low-dimensional features of the target.

[0009] The sparse features are input into a feature learning model, which then obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features.

[0010] The category semantic information is input into the classification model to identify targets in the SAR image.

[0011] In one embodiment, the convolutional sparse coding model based on a deep unfolded network is constructed based on a learnable soft thresholding iterative algorithm, including multiple sequentially connected feature extraction units, each of which corresponds to one iteration calculation.

[0012] In one embodiment, each of the feature extraction units performs the following calculations:

[0013]

[0014] in,

[0015] In the above formula, Let represent the non-negative soft thresholding shrinkage operator, θ and 1 / L represent the threshold and matrix factor, respectively, both of which are learnable variable values, D represents the learnable dictionary parameters, X represents the SAR image input to the first feature extraction unit, and Z... i Z represents the feature extracted by the i-th feature extraction unit. i+1 This represents the features extracted by the (i+1)th feature extraction unit. Conv(·) represents the convolution operation, and Deconv(·) represents the deconvolution operation.

[0016] In one embodiment, the feature learning model includes a pooling layer, a residual layer, and a global average pooling layer connected in sequence.

[0017] In one embodiment, the step of inputting the sparse features into a feature learning model, whereby the feature learning model obtains discriminative feature representations from the sparse features and extracts latent category semantic features from the sparse features includes:

[0018] The size of the sparse feature is compressed using a pooling layer with a step size of 2 to obtain the compressed sparse feature.

[0019] The compressed sparse features are extracted using four sequentially connected residual layers to obtain the high-dimensional semantic information of the target.

[0020] The dimensionality of the high-dimensional semantic information is reduced by using the global average pooling layer to obtain the category semantic features.

[0021] In one embodiment, the convolutional sparse coding model, feature learning model, and classification model based on the deep unfolded network are sequentially connected to construct an interpretable target recognition network.

[0022] The SAR image to be identified is input into the interpretable target identification network to perform interpretable target identification in the SAR image.

[0023] In one embodiment, the target recognition network is trained using a cross-entropy loss function, which is expressed as:

[0024]

[0025] In the above formula, N represents the number of training samples, C represents the number of classes, and y ic The sign function p is 1 if and only if the true class of sample i is c, otherwise it is 0. ic This represents the probability that observed sample i belongs to category c, which can be obtained by predicting the label and applying the Softmax function.

[0026] An interpretable target recognition device based on a deep sparse coding model, the device comprising:

[0027] The SAR image acquisition module is used to acquire SAR images for target identification.

[0028] The sparse feature extraction module is used to input the SAR image into a convolutional sparse coding model based on a deep unfolding network, and the model extracts sparse features representing low-dimensional features of the target after sparse decomposition of the SAR image.

[0029] The category semantic feature extraction module is used to input the sparse features into the feature learning model, which then obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features.

[0030] The target recognition module is used to input the category semantic information into the classification model to identify targets in the SAR image.

[0031] 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:

[0032] Acquire SAR images for target identification;

[0033] The SAR image is input into a convolutional sparse coding model based on a deep unfolding network. The model performs sparse decomposition on the SAR image and extracts sparse features representing low-dimensional features of the target.

[0034] The sparse features are input into a feature learning model, which then obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features.

[0035] The category semantic information is input into the classification model to identify targets in the SAR image.

[0036] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0037] Acquire SAR images for target identification;

[0038] The SAR image is input into a convolutional sparse coding model based on a deep unfolding network. The model performs sparse decomposition on the SAR image and extracts sparse features representing low-dimensional features of the target.

[0039] The sparse features are input into a feature learning model, which then obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features.

[0040] The category semantic information is input into the classification model to identify targets in the SAR image.

[0041] The aforementioned interpretable target recognition method, apparatus, and device based on a deep sparse coding model first performs sparse decomposition on the SAR image to be identified using a convolutional sparse coding model based on a deep unfolded network to extract sparse features representing low-dimensional target features. Then, a feature learning model is used to obtain discriminative feature representations from the sparse features and extract latent category semantic features from them. Finally, a classification model identifies the target in the SAR image. This method achieves accurate target classification and recognition in SAR images while enhancing the interpretability of the model during image processing. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating an interpretable target recognition method based on a deep sparse coding model in one embodiment.

[0043] Figure 2 This is a schematic diagram of the structural framework of an interpretable target recognition network in one embodiment;

[0044] Figure 3 This is a schematic diagram of the feature extraction unit in one embodiment;

[0045] Figure 4 This is a structural block diagram of an interpretable target recognition device based on a deep sparse coding model in one embodiment;

[0046] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0047] 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.

[0048] In existing technologies, when using deep learning methods to identify targets in SAR images, the lack of deep integration of expert knowledge / theoretical models with deep network learning still fails to effectively improve model interpretability. Theoretical models possess strong interpretability; constructing a hybrid model with deep learning methods can enhance the algorithm's transparency and interpretability while effectively utilizing data. To address the problem of insufficient interpretability and lack of theoretical model support in existing deep learning model designs, in one embodiment, such as... Figure 1 As shown, an interpretable target recognition method based on a deep sparse coding model is provided, including the following steps:

[0049] Step S100: Acquire the SAR image to be identified;

[0050] Step S110: Input the SAR image into the convolutional sparse coding model based on the deep unfolding network. The model performs sparse decomposition on the SAR image and extracts sparse features representing the low-dimensional features of the target.

[0051] In step S120, sparse features are input into a feature learning model, which then obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features.

[0052] In step S130, the category semantic information is input into the classification model to identify targets in the SAR image.

[0053] In this embodiment, a convolutional sparse coding model is first used to sparsely encode targets in SAR images, extracting sparse features for reconstruction. To better integrate into the end-to-end learning framework, the convolutional sparse coding model is solved using a deep unfolded network designed based on sparse reconstruction criteria. Then, guided by a classification task, a feature learning model is used to obtain discriminative feature representations from the sparse features. Finally, a classifier is used to perform classification prediction based on the features obtained from the feature learning network. Within the end-to-end classification framework, this method can learn sparse features with both target reconstruction and discriminative attributes from the solution space of the convolutional sparse coding model, and can construct a direct mapping between the sparse feature space and the SAR image space, exhibiting good pre-interpretability. Furthermore, this method can improve the target recognition rate when performing target identification on SAR images obtained under different elevation angle conditions.

[0054] In this embodiment, the convolutional sparse coding model, feature learning model, and classification model based on a deep unfolded network can be sequentially connected to construct an interpretable target recognition network, which is the framework model for target recognition in this method, such as... Figure 2 As shown, unlike the traditional top-down deep learning framework of CNN models, the target recognition model in this method introduces a convolutional sparse coding model into the deep learning network framework. It generates potential low-level feature representations through a bottom-up architecture, and then fully utilizes a top-down feature learning network structure to learn discriminative feature representations. On one hand, this model combines the traditional CNN model with the convolutional sparse coding model, linking the SAR target recognition task with the low-level sparse feature learning task, and using the convolutional sparse coding model to obtain discriminative low-level feature representations. On the other hand, for the overall model, the convolutional sparse coding model not only constructs a direct mapping relationship between the sparse feature space and the image space, but also has the ability to represent key discriminative information, improving the interpretability of the overall model.

[0055] In step S100, the SAR image to be identified is acquired. In order to make better use of the inherent sparsity of SAR, the SAR image data without any preprocessing is directly acquired from the echo data.

[0056] In step S110, the SAR image is first input into the convolutional sparse coding model based on a deep unfolded network in the interpretable target recognition network. This model is constructed based on a learnable soft thresholding iterative algorithm and includes multiple sequentially connected feature extraction units, with each feature extraction unit corresponding to one iteration calculation.

[0057] Specifically, in interpretable target recognition networks, convolutional sparse coding models are used to model targets in SAR images and extract sparse features for reconstruction. These sparse features can be deconvolved with corresponding dictionaries to obtain image reconstruction components, thus building a bridge between features and target reconstruction components, leading to more intuitive interpretability. However, the sparse solution space of convolutional sparse coding models is infinite. How to extract discriminative sparse features from the solution space is a more pressing issue for classification tasks. Guided by classification tasks, convolutional sparse coding models learn sparse features that simultaneously possess the attributes of target reconstruction and discriminative analysis, which forms the foundation for the interpretability of interpretable target recognition networks.

[0058] However, convolutional sparse coding algorithms, such as ADMM, are often limited by high computational resource requirements and low computational efficiency, making them difficult to use for large-scale data training and real-time systems.

[0059] In this embodiment, the convolutional sparse coding model based on a deep unfolded network can efficiently perform sparse coding solution and training in a task-driven manner within an end-to-end framework. In this method, the convolutional sparse coding model uses convolution to reconstruct the image; therefore, the feature extraction units used to solve the convolutional sparse coding model, such as... Figure 3 As shown, the specific manifestation is as follows:

[0060]

[0061] In formula (1), Let represent the non-negative soft thresholding shrinkage operator, θ and 1 / L represent the threshold and matrix factor, respectively, both of which are learnable variable values, D represents the learnable dictionary parameters, X represents the SAR image input to the first feature extraction unit, and Z... i Z represents the feature extracted by the i-th feature extraction unit. i+1 This represents the features extracted by the (i+1)th feature extraction unit. Conv(·) represents the convolution operation, and Deconv(·) represents the deconvolution operation.

[0062] In convolutional sparse coding models based on deep unfolded networks, several such... Figure 3 The cascading of feature extraction units shown is used to efficiently obtain approximate estimates of the sparse coding of the signal.

[0063] In step S120, after the convolutional sparse coding model performs sparse decomposition on the target in the SAR image and obtains its low-level sparse feature representation, for classification tasks, although the learned sparse features have a certain degree of discriminativeness, there is still a lack of effective utilization of sparse features. Traditional CNN models use multi-layer convolutional structures to obtain discriminative features from images, showing outstanding advantages in extracting high-dimensional semantic information. However, in interpretable target recognition networks, feature learning networks are used to obtain discriminative feature representations from sparse features and extract the potential category semantic information of sparse features.

[0064] In this embodiment, the feature learning model includes a pooling layer, a residual layer, and a global average pooling layer connected in sequence. To fully utilize the feature learning capabilities of deep network structures, firstly, to reduce the number of network parameters, the feature learning network uses a pooling layer with a stride of 2 to compress the size of sparse features. Secondly, since sparse features often exhibit strong sparsity, even having only a few pixels, directly using convolution operations may cause the loss of key information represented by sparse features. Therefore, the feature network uses four basic residual modules to extract high-dimensional semantic information of the target. Another reason for using residual modules is to avoid gradient loss of sparse features in deep network structures. Then, the global average pooling layer is used to further reduce the feature dimensionality and generate a discriminative feature representation of a feature vector with a dimension of 512, which is the category semantic feature.

[0065] In this embodiment, the structure of the feature learning model is shown in Table 1.

[0066] Table 1 Feature Learning Network Structure

[0067]

[0068] Finally, in step S130, the classifier predicts and classifies the extracted category semantic information. The classifier can be, for example, a linear classifier, an SVM (Support Vector Basis), or a decision tree.

[0069] In this embodiment, when training the interpretable target recognition network, the parameters in the network are updated based on the backpropagation of the classification loss. Specifically, the cross-entropy loss function is used for training. The cross-entropy loss function is calculated using the true labels of the training samples and the predicted labels obtained by the interpretable target recognition network. The cross-entropy loss function is expressed as follows:

[0070]

[0071] In formula (2), N represents the number of training samples, C represents the number of classes, and y ic The sign function p is 1 if and only if the true class of sample i is c, otherwise it is 0. ic This represents the probability that observed sample i belongs to category c, which can be obtained by predicting the label and applying the Softmax function.

[0072] In step S140, the pre-interpretability of the model is explored by using the input samples, reconstructed samples, sparse feature maps, and corresponding reconstructed component maps obtained from the trained model. Furthermore, the class activation maps of specific modules of the convolutional sparse coding model in the model are visualized using LayerCAM to explore the post-interpretability of the model.

[0073] In this embodiment, we first explore the model's pre-interpretability. The model can reconstruct the input image, and by comparing the reconstructed image of the convolutional sparse coding model with the feature representation learned from the model's input samples, the model is more interpretable. This model can achieve a direct mapping between the sparse feature space and the image space, possessing more intuitive interpretability. Next, we explore the model's post-interpretability. Based on LayerCAM, we visualize the class activation maps of specific modules in the convolutional sparse coding model, using feature map gradients to obtain the importance of local locations to the class activation maps. For the same sequence of sample class activation maps, the model's class activation maps show less variation, and the decision regions are relatively more stable, making its interpretability more reliable compared to traditional CNN models.

[0074] In this embodiment, SAR images with an elevation angle of 17 degrees are used as the training set. By conducting comparative experiments on test sets with different elevation angles using the method in this paper and some classical methods, it can be demonstrated that the method can improve the target recognition rate when performing target recognition on SAR images under different elevation angle conditions.

[0075] Classical methods employed include A-ConvNet, JSRC, ASC_Match, FEC, and ResNet18. The experiment also compared the recognition performance of traditional CNN and DCSC models, with results shown in Table 2. The typical CNN and DCSC models have the same number of channels in the first layer, 64 channels in total. The main difference lies in the number of unfolded layers in the convolutional sparse coding model. When the number of unfolded layers in the convolutional sparse coding model is 0, the DCSC model degenerates into a typical CNN model.

[0076] Table 2 Comparison of experimental results of various methods at different pitch angles.

[0077]

[0078] The experimental results lead to the following conclusions: ① As the difference in imaging elevation angle between the training and test sets increases, the recognition rates of different algorithms all decrease. For experiment EOC-15, due to the small difference in imaging elevation angle between the training and test data, the recognition rates of all algorithms can reach over 98%. However, for experiment EOC-45, the scattering characteristics between the training and test data change significantly, resulting in a marked decrease in the recognition accuracy of all methods. ② In terms of average recognition rate, the DCSC model (the interpretable target recognition network proposed in this paper) outperforms the comparative experimental methods, verifying the effectiveness of this method.

[0079] Compared to other comparative methods, the main advantage of this method lies in EOC-45 recognition, that is, under the condition of large pitch angle changes, the DCSC model still maintains relatively high recognition performance.

[0080] In addressing the shortcomings of existing deep learning-based SAR target recognition models, which lack theoretical support and interpretability, this paper proposes a new interpretable SAR target recognition method based on a deep convolutional sparse coding model. Within a classification task-driven framework, the model first guides a deep unrolled convolutional sparse coding model to learn sparse features that simultaneously possess target reconstruction and discriminative attributes, making the network structure pre-interpretable. Then, a feature learning network extracts discriminative feature representations from the sparse features for classification. Finally, the classification loss is calculated based on the predicted and ground truth labels of the target, and the model parameters are updated through backpropagation of the gradient of the classification loss. This target recognition method maintains good classification performance while extracting sparse features corresponding to strong scattering points and basic scattering structures. By convolving a convolutional dictionary with the sparse feature map, the reconstructed components of the target are obtained, allowing for a direct mapping between the sparse feature space and the image space, resulting in more intuitive interpretability. Within a certain range, the number of convolutional channels and the number of depthwise unwrap layers in the convolutional sparse coding model of this method do not significantly affect the classification performance of the model. However, as the number of convolutional channels increases, the model learns richer features and enhances its ability to describe target details, but redundancy also increases. As the number of depthwise unwrap layers increases, the sparsity of the learned features increases, redundancy decreases, and the interpretability of the model is enhanced. Compared with traditional CNN models, the class activation map of this method focuses more on the strong scattering structure of the target, effectively suppressing background information and possessing more reliable interpretability. Furthermore, experiments demonstrate that this method still maintains a high recognition rate when the pitch angle varies significantly.

[0081] It should be understood that, although Figure 1The 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 1 At 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.

[0082] In one embodiment, such as Figure 4 As shown, an interpretable target recognition device based on a deep sparse coding model is provided, comprising: a SAR image acquisition module 200, a sparse feature extraction module 210, a category semantic feature extraction module 220, and a target recognition module 230, wherein:

[0083] SAR image acquisition module 200 is used to acquire SAR images for target identification.

[0084] The sparse feature extraction module 210 is used to input the SAR image into a convolutional sparse coding model based on a deep unfolding network, and extract sparse features representing low-dimensional features of the target after the SAR image is sparsely decomposed by the model.

[0085] The category semantic feature extraction module 220 is used to input the sparse features into the feature learning model, and the feature learning model obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features;

[0086] The target recognition module 230 is used to input the category semantic information into the classification model to recognize targets in the SAR image.

[0087] Specific limitations regarding the interpretable target recognition device based on the deep sparse coding model can be found in the limitations of the interpretable target recognition method based on the deep sparse coding model mentioned above, and will not be repeated here. Each module in the aforementioned interpretable target recognition device based on the deep sparse coding model 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, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0088] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements an interpretable target recognition method based on a deep sparse coding model. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0089] Those skilled in the art will understand that Figure 5 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.

[0090] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0091] Acquire SAR images for target identification;

[0092] The SAR image is input into a convolutional sparse coding model based on a deep unfolding network. The model performs sparse decomposition on the SAR image and extracts sparse features representing low-dimensional features of the target.

[0093] The sparse features are input into a feature learning model, which then obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features.

[0094] The category semantic information is input into the classification model to identify targets in the SAR image.

[0095] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0096] Acquire SAR images for target identification;

[0097] The SAR image is input into a convolutional sparse coding model based on a deep unfolding network. The model performs sparse decomposition on the SAR image and extracts sparse features representing low-dimensional features of the target.

[0098] The sparse features are input into a feature learning model, which then obtains discriminative feature representations from the sparse features and extracts potential category semantic features from the sparse features.

[0099] The category semantic information is input into the classification model to identify targets in the SAR image.

[0100] 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.

[0101] 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.

[0102] 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. An interpretable object recognition method based on a deep sparse coding model, characterized in that, The method comprises: acquiring a SAR image to be subjected to target recognition; inputting the SAR image into a convolutional sparse coding model based on a deep unfolding network, extracting sparse features representing low-dimensional features of a target after the SAR image is subjected to sparse decomposition by the model; inputting the sparse features into a feature learning model, acquiring discriminative feature representation from the sparse features by the feature learning model, and extracting latent class semantic features in the sparse features, in the feature learning model, compressing the size of the sparse features by a pooling layer with a step size of 2 to obtain compressed sparse features, extracting features of the compressed sparse features by four residual layers connected in sequence to obtain high-dimensional semantic information of the target, and reducing the dimension of the high-dimensional semantic information by a global average pooling layer to obtain the class semantic features; inputting the class semantic features into a classification model to recognize the target in the SAR image. 2.The explainable target identification method of claim 1, wherein, The convolutional sparse coding model based on the deep unfolding network is constructed based on a learnable soft threshold iterative algorithm, and comprises a plurality of feature extraction units connected in sequence, each of the feature extraction units corresponding to one iteration calculation.

3. The interpretable object recognition method of claim 2, wherein, Each of the feature extraction units performs the following calculation: In the above formula, denotes a non-negative soft threshold shrinkage operator, and denote threshold and matrix factor, respectively, both are learnable variable values, denotes a learnable dictionary parameter, denotes the input SAR image to the first feature extraction unit, denotes the feature extracted by the first feature extraction unit, denotes the feature extracted by the first feature extraction unit, denotes a convolution operation, denotes a deconvolution operation. 4.The interpretable object recognition method of claim 1, wherein, The feature learning model comprises a pooling layer, a residual layer, and a global average pooling layer connected in sequence.

5. The interpretable object recognition method of any one of claims 1-4, wherein, The convolutional sparse coding model based on the deep unfolding network, the feature learning model, and the classification model are connected in sequence to construct an interpretable target recognition network; inputting the SAR image to be subjected to target recognition into the interpretable target recognition network to perform interpretable recognition on the target in the SAR image.

6. The interpretable object recognition method of claim 5, wherein, When the target recognition network is trained, a cross-entropy loss function is used to train the target recognition network, and the cross-entropy loss function is expressed as: In the above formula, denotes the number of training samples, denotes the number of classes, denotes the symbol function, and takes the value 1 if and only if the sample has the true class , and 0 otherwise, denotes the probability that the observed sample belongs to the class , which is obtained by passing the predicted label through a Softmax function.

7. An interpretable object recognition apparatus based on a deep sparse coding model, characterized by, The device comprises: a SAR image acquisition module configured to acquire a SAR image to be subjected to target recognition; a sparse feature extraction module configured to input the SAR image into a convolutional sparse coding model based on a deep unfolding network, extract sparse features representing low-dimensional features of a target after the SAR image is subjected to sparse decomposition by the model, compress the size of the sparse features by a pooling layer with a step size of 2 in a feature learning model to obtain compressed sparse features, extract features of the compressed sparse features by four residual layers connected in sequence to obtain high-dimensional semantic information of the target, and reduce the dimension of the high-dimensional semantic information by a global average pooling layer to obtain class semantic features; a class semantic feature extraction module configured to input the sparse features into the feature learning model, acquire discriminative feature representation from the sparse features by the feature learning model, and extract latent class semantic features in the sparse features; a target recognition module configured to input the class semantic features into a classification model to recognize the target in the SAR image.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor implements the steps of the method of any one of claims 1 to 6 when executing the computer program.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 6.

Images