Radar target sar image joint detection and recognition method based on saliency migration

Abstract

The application relates to a radar target SAR image joint detection and recognition method based on saliency migration. The method comprises the following steps: pre-processing a radar target two-dimensional image to obtain a SAR grayscale image; dividing the SAR grayscale image into a training set and a test set, performing random data enhancement and normalization processing on the training set to obtain a normalized training set; constructing a deep learning model; the deep learning model comprises a double-layer convolution, a capsule network, saliency detection and a feature layer mask; training the deep learning model by using the normalized training set to obtain a trained deep learning model; and identifying the test set according to the trained deep learning model to obtain an identification result. The method can improve the detection accuracy of a radar target image.

Description

Technical Field

[0001] This application relates to the field of radar target recognition technology, and in particular to a method, apparatus, computer equipment, and storage medium for joint detection and recognition of radar target SAR images based on saliency migration. Background Technology

[0002] Synthetic space radar (SAR), as an active imaging device, has significant application value in remote sensing. Because it can operate in all weather conditions and around the clock, it has been widely used in military and civilian fields, such as automatic target identification, battlefield situational awareness, and Earth remote sensing and control. Due to the unique imaging mechanism of SAR, SAR images contain both target and background clutter. Therefore, a fundamental problem in the field of automatic target identification using SAR images is detecting and identifying targets within background clutter.

[0003] Existing methods typically decompose this problem into two stages: a detection task to separate the target from background clutter in SAR images, and an identification task to determine the target type. Currently, deep learning techniques widely used in automatic SAR target identification often neglect or simplify the detection task, employing methods such as cropping the target region or threshold filtering to remove background clutter. Cropping preserves background clutter within the slice, while threshold filtering relies on parameter selection and does not integrate detection and identification tasks, resulting in low accuracy in target detection and identification across different scenarios. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, and storage medium for joint detection and recognition of radar target SAR images based on saliency transfer, which can improve the accuracy of radar target SAR image detection and recognition in different scenarios, in order to address the above-mentioned technical problems.

[0005] A joint detection and recognition method for radar target SAR images based on saliency migration, the method comprising:

[0006] Acquire a two-dimensional image of the radar target to be identified;

[0007] Preprocessing is performed on the two-dimensional image of the radar target to obtain a SAR grayscale image;

[0008] The SAR grayscale images are divided into training and test sets. The training set is then subjected to random data augmentation and normalization to obtain the normalized training set.

[0009] Construct a deep learning model; the deep learning model includes two-layer convolution, capsule network, saliency detection, and feature layer masking; the deep learning model has a U-shaped structure;

[0010] The deep learning model is trained using the normalized training set to obtain a well-trained deep learning model.

[0011] The test set is identified using a trained deep learning model to obtain the identification results.

[0012] In one embodiment, the deep learning model is trained using a normalized training set to obtain a trained deep learning model, including:

[0013] Obtain multiple different training set samples from the normalized training set;

[0014] By using double-layer convolution to extract target information from samples in different training sets, multi-scale target features are obtained.

[0015] Based on the U-shaped structure of the deep learning model, feature aggregation is performed on multi-scale target features to obtain enhanced multi-scale target features;

[0016] The target importance information is transferred from the optical pre-trained model by using the saliency map as a pseudo-label for the saliency detection to obtain the target key region;

[0017] Based on the feature layer mask, target and background separation is performed on the enhanced multi-scale target features to obtain the target features;

[0018] The target features are input into the capsule network for feature transformation to obtain the target spatial features;

[0019] The deep learning model is trained by using the target key region, target spatial features and a pre-set loss function to obtain a trained deep learning model.

[0020] In one embodiment, the process of obtaining a saliency map includes:

[0021] The VGG16 model pre-trained on ImageNet was trained on the standard working condition dataset of MSTAR to obtain a trained VGG16 model.

[0022] Various saliency maps are extracted from the trained VGG16 model using Guided Grad-Cam.

[0023] In one embodiment, the pre-defined loss function includes the capsule network's margin loss, the saliency detection binary classification cross-entropy loss, and the target region mask L1 regularization.

[0024] In one embodiment, the pre-set loss function is:

[0025] L = L m +α·L S +β·L1

[0026] L j =Tj max(0,m + -||v j ||) 2 +

[0027] λ(1-T j max(0,||v) j ||-m - ) 2

[0028]

[0029] Where, m + =0.9 represents the lower bound for true positives, m - =0.1 represents the upper bound of the true negative, λ=0.5 represents the proportionality coefficient, L m L represents the interval loss of the capsule network. S L1 represents the cross-entropy loss for saliency detection binary classification, and L1 represents the 1-norm regularization of the target region mask. j Let T represent the loss of class j. j T represents the indicator function. j =1 is only for the j-th class, with hyperparameters α set to 1e-5 and β set to 5e-6 to control the sparsity of the mask.

[0030] In one embodiment, the two-dimensional image of the radar target is preprocessed to obtain a SAR grayscale image, including:

[0031] A linear transformation and grayscale enhancement are performed on the two-dimensional image of the radar target to obtain a SAR grayscale image;

[0032] The calculation process of linear transformation is as follows:

[0033]

[0034] Where x represents the original radar target 2D image pixel input value, and y represents the output image pixel value, x max x represents the maximum pixel value of the original two-dimensional radar target image. min This represents the minimum pixel value in the original two-dimensional radar target image;

[0035] The calculation process for grayscale enhancement is as follows:

[0036]

[0037] Where I represents the grayscale value of the radar target 2D image, and O represents the grayscale value of the SAR grayscale image. min_count I represents the grayscale value that appears least frequently in a two-dimensional radar target image. max_count This represents the grayscale value that appears most frequently in a two-dimensional radar target image.

[0038] In one embodiment, the training set is subjected to random data augmentation and normalization to obtain a normalized training set, including:

[0039] The training set is subjected to random data augmentation and normalization based on image rotation, Gaussian white noise perturbation, random uniform noise replacement, and a pre-set random data augmentation method implementation rate, resulting in a normalized training set; the implementation rates of each method in the random data augmentation are 0.3:0.2:0.2, respectively.

[0040] A radar target SAR image joint detection and recognition device based on saliency migration, the device comprising:

[0041] The preprocessing module is used to acquire a two-dimensional image of the radar target to be identified; and to preprocess the two-dimensional image of the radar target to obtain a SAR grayscale image.

[0042] The random augmentation and normalization module is used to divide SAR grayscale images into training and test sets, and to perform random data augmentation and normalization on the training set to obtain the normalized training set.

[0043] The model building and training module is used to build deep learning models; the deep learning models include two-layer convolution, capsule networks, saliency detection, and feature layer masks; the deep learning models have a U-shaped structure; the deep learning models are trained using normalized training sets to obtain trained deep learning models;

[0044] The recognition module is used to identify the test set based on the trained deep learning model and obtain the recognition results.

[0045] 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:

[0046] Acquire a two-dimensional image of the radar target to be identified;

[0047] Preprocessing is performed on the two-dimensional image of the radar target to obtain a SAR grayscale image;

[0048] The SAR grayscale images are divided into training and test sets. The training set is then subjected to random data augmentation and normalization to obtain the normalized training set.

[0049] Construct a deep learning model; the deep learning model includes two-layer convolution, capsule network, saliency detection, and feature layer masking; the deep learning model has a U-shaped structure;

[0050] The deep learning model is trained using the normalized training set to obtain a well-trained deep learning model.

[0051] The test set is identified using a trained deep learning model to obtain the identification results.

[0052] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0053] Acquire a two-dimensional image of the radar target to be identified;

[0054] Preprocessing is performed on the two-dimensional image of the radar target to obtain a SAR grayscale image;

[0055] The SAR grayscale images are divided into training and test sets. The training set is then subjected to random data augmentation and normalization to obtain the normalized training set.

[0056] Construct a deep learning model; the deep learning model includes two-layer convolution, capsule network, saliency detection, and feature layer masking; the deep learning model has a U-shaped structure;

[0057] The deep learning model is trained using the normalized training set to obtain a well-trained deep learning model.

[0058] The test set is identified using a trained deep learning model to obtain the identification results.

[0059] The aforementioned method, apparatus, computer equipment, and storage medium for joint detection and recognition of radar targets based on saliency transfer in SAR images first preprocesses the two-dimensional radar target image to obtain a SAR grayscale image, enhancing the target details in the SAR image and enabling more accurate extraction of image feature information, thereby improving the accuracy of image detection. The SAR grayscale image is divided into a training set and a test set. Random data augmentation and normalization are performed on the training set, and a deep learning model is constructed from the normalized training set. The deep learning model includes two-layer convolution, capsule networks, saliency detection, and feature layer masking. The saliency map of the pre-trained model is used as the target image. The pseudo-labels for saliency detection transfer information about key feature regions of the target. Feature layer masks are used to separate the target from the background. By integrating saliency detection and capsule network recognition into a deep learning model as detection and recognition functional modules, the deep learning model can effectively extract spatial information of the target in the scene, improve the accuracy of detection and recognition, and has good generalization performance. It also has good recognition robustness in different scenarios and can detect the location of targets in unknown scenarios. Furthermore, the saliency map of the pre-trained model is used as a pseudo-label for saliency detection to transfer the importance information of the target, avoiding the time-consuming and costly drawbacks of manual annotation. Attached Figure Description

[0060] Figure 1This is a flowchart illustrating a joint detection and recognition method for radar targets based on saliency migration in one embodiment;

[0061] Figure 2 Here is a diagram of the deep learning model structure in one embodiment;

[0062] Figure 3 This is a pseudo-label map of a saliency map used for a saliency detection task in one embodiment;

[0063] Figure 4 This is a schematic diagram of the MSTAR dataset detection results in different scenarios in another embodiment;

[0064] Figure 5 This is a schematic diagram of the detection structure for ships on the sea surface under different scenarios in another embodiment;

[0065] Figure 6 This is a structural block diagram of a radar target SAR image joint detection and recognition device based on saliency migration in one embodiment;

[0066] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0067] 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.

[0068] In one embodiment, such as Figure 1 As shown, a joint detection and recognition method for radar targets based on saliency migration SAR images is provided, including the following steps:

[0069] Step 102: Obtain a two-dimensional image of the radar target to be identified; preprocess the two-dimensional image of the radar target to obtain a SAR grayscale image.

[0070] The original synthetic aperture radar (SAR) 2D images of radar targets are transformed into SAR grayscale images through data preprocessing methods such as linear transformation and grayscale enhancement. This enhances the target details in the SAR images, enabling more accurate extraction of image feature information and thus improving the accuracy of image detection. The processed SAR grayscale images are divided into training and testing sets for subsequent model training. Each SAR grayscale image is a 2D matrix.

[0071] Step 104: Divide the SAR grayscale image into a training set and a test set, and perform random data augmentation and normalization on the training set to obtain the normalized training set.

[0072] Random data augmentation of the training set enhances sample diversity. After normalization, these samples are used as training samples, improving model training accuracy. Random data augmentation methods include image rotation, additive white Gaussian noise (ADG), and uniform noise replacement. The probabilities of applying image rotation, ADG, and uniform noise are 0.3, 0.2, and 0.2, respectively. Image rotation rotates the original image by a small angle, enhancing sample diversity; the rotation angle is uniformly distributed between -5 and 5. Additive white Gaussian noise (ADG) adds ADG to the original image to simulate the impact of different scene changes on the intensity of target scattering points; its mean is 0.1, standard deviation is 0.1, and amplitude is uniformly distributed between 0.5 and 1.5. Uniform noise replacement uses uniformly distributed noise between 0 and 1, replacing a portion of the original image between 0 and 0.05 to avoid the model's sensitivity to changes in single pixels.

[0073] Step 106: Construct a deep learning model; the deep learning model includes two-layer convolution, capsule network, saliency detection, and feature layer masking; the deep learning model has a U-shaped structure; train the deep learning model using the normalized training set to obtain a trained deep learning model.

[0074] The deep learning model in this application is designed as a U-shaped structure. The information of each of the first four layers of the U-shaped structure is used as multi-scale feature information for feature aggregation. The feature information of different scales in each layer describes the structural information of the target at different levels. The low-level feature information represents the target texture, structure and other information, while the high-level feature information represents the target semantic category and other information. This enhances the performance of the detection and recognition function module of the integrated design of saliency detection and capsule network recognition.

[0075] The deep learning model includes two-layer convolution, capsule networks, saliency detection, and feature layer masks. It utilizes the saliency map of the pre-trained model as pseudo-labels for saliency detection to transfer information about key feature regions of the target. Feature layer masks are used to separate the target from the background. By integrating saliency detection and capsule network recognition into the deep learning model as detection and recognition modules, the model can effectively extract spatial information of the target in the scene, improving detection and recognition accuracy. It also exhibits good generalization performance, strong robustness in different scenarios, and can detect targets in unknown scenes. Furthermore, using the saliency map of the pre-trained model as pseudo-labels for saliency detection transfers target importance information, avoiding the time-consuming and costly nature of manual annotation. The pseudo-label map is shown below. Figure 2 As shown, the first row contains the SAR grayscale images from the MSTAR dataset, and the second row contains the corresponding saliency map results.

[0076] Step 108: Use the trained deep learning model to identify the test set and obtain the identification results.

[0077] In the aforementioned joint detection and recognition method for radar targets based on saliency transfer in SAR images, the two-dimensional image of the radar target is first preprocessed to obtain a SAR grayscale image, which enhances the target details in the SAR image, enabling more accurate extraction of image feature information and thus improving the accuracy of image detection. The SAR grayscale image is divided into a training set and a test set. The training set undergoes random data augmentation and normalization processing. The normalized training set is used to construct a deep learning model. The deep learning model includes two-layer convolution, capsule network, saliency detection, and feature layer masking. The saliency map of the pre-trained model is used as the saliency detection... The pseudo-labels transferred key feature regions of the target, and the feature layer mask was used to separate the target from the background. By integrating saliency detection and capsule network recognition into the deep learning model as detection and recognition functional modules, the deep learning model can effectively extract the spatial information of the target in the scene, improve the accuracy of detection and recognition, and has good generalization performance. It has good recognition robustness in different scenes and can also detect the position of the target in unknown scenes. Furthermore, the saliency map of the pre-trained model is used as the pseudo-label for saliency detection to transfer the importance information of the target, avoiding the time-consuming and costly drawbacks of manual annotation.

[0078] In one embodiment, the deep learning model is trained using a normalized training set to obtain a trained deep learning model, including:

[0079] Obtain multiple different training set samples from the normalized training set;

[0080] By using double-layer convolution to extract target information from samples in different training sets, multi-scale target features are obtained.

[0081] Based on the U-shaped structure of the deep learning model, feature aggregation is performed on multi-scale target features to obtain enhanced multi-scale target features;

[0082] The target importance information is transferred from the optical pre-trained model by using the saliency map as a pseudo-label for the saliency detection to obtain the target key region;

[0083] Based on the feature layer mask, target and background separation is performed on the enhanced multi-scale target features to obtain the target features;

[0084] The target features are input into the capsule network for feature transformation to obtain the target spatial features;

[0085] The deep learning model is trained by using the target key region, target spatial features and a pre-set loss function to obtain a trained deep learning model.

[0086] In one embodiment, the process of obtaining a saliency map includes:

[0087] The VGG16 model pre-trained on ImageNet was trained on the standard working condition dataset of MSTAR to obtain a trained VGG16 model.

[0088] Various saliency maps are extracted from the trained VGG16 model using Guided Grad-Cam.

[0089] In a specific embodiment, the various saliency maps are summed to eliminate class distinctions and normalized to [0, 1] to obtain the target key region.

[0090] In one embodiment, the pre-defined loss function includes the capsule network's margin loss, the saliency detection binary classification cross-entropy loss, and the target region mask L1 regularization.

[0091] In one embodiment, the pre-set loss function is:

[0092] L = L m +α·L S +β·L1

[0093] L j =T j max(0,m + -||v j ||) 2 +

[0094] λ(1-T j max(0,||v) j ||-m - ) 2

[0095]

[0096] Where, m + =0.9 represents the lower bound for true positives, m - =0.1 represents the upper bound of the true negative, λ=0.5 represents the proportionality coefficient, L m L represents the interval loss of the capsule network. S L1 represents the cross-entropy loss for saliency detection binary classification, and L1 represents the 1-norm regularization of the target region mask. j Let T represent the loss of class j. j T represents the indicator function. j =1 is only for the j-th class, with hyperparameters α set to 1e-5 and β set to 5e-6 to control the sparsity of the mask.

[0097] In a specific embodiment, a deep learning model combining U-shaped structure, saliency detection, and feature layer masking is used, such as... Figure 2 As shown, the adaptive average pooling layer ensures that the input image size is not fixed at 128×128, but adapts to different image sizes. The double convolutional layer uses two convolutional layers, a batch normalization layer structure, and the ReLU activation function. The kernel size of the convolutional layer is 5×5, the stride is 1×1, and the padding size is 2×2. Figure 2 The U-shaped structure shown utilizes information at different scales to enhance the performance of the detection and recognition modules. Specifically, the information at different scales is obtained by using information from each of the first four segments of the U-shaped structure as multi-scale feature information.

[0098] The Guided Grad-Cam was used to extract saliency maps from the VGG16 pre-trained model. The VGG16 model, pre-trained on ImageNet, was trained on the MSTAR Standard Operating Condition (SOC) dataset. The saliency maps of the trained VGG16 model were extracted using the Guided Grad-Cam, and various saliency maps were mapped. c Summing yields a Map multi To eliminate class distinctions and normalize to [0, 1], the summation of the saliency maps of each class is used as pseudo-labels to eliminate class distinctions and normalize to [0, 1]. Feature layer masking is performed in the intermediate feature layer to utilize the good generalization performance of the intermediate layer features, resulting in good detection generalization for unknown scenes and targets. The recognition task uses a capsule network to process multi-scale features to obtain the recognition results. The initial capsule layer convolution kernel size of the capsule network is 5×5, and the stride is 2×2. The input capsule size of the routing capsule layer is 576×8, and the output capsule length is 8.

[0099] During model training, the network training parameters are as follows: training batch size is 100, batch size is 64, learning rate is 3e-4, exponential decay rate is 0.98, and the optimizer uses the NAdam algorithm.

[0100] In one embodiment, the two-dimensional image of the radar target is preprocessed to obtain a SAR grayscale image, including:

[0101] A linear transformation and grayscale enhancement are performed on the two-dimensional image of the radar target to obtain a SAR grayscale image;

[0102] The calculation process of linear transformation is as follows:

[0103]

[0104] Where x represents the original radar target 2D image pixel input value, and y represents the output image pixel value, x max x represents the maximum pixel value of the original two-dimensional radar target image. minThis represents the minimum pixel value in the original two-dimensional radar target image;

[0105] The calculation process for grayscale enhancement is as follows:

[0106]

[0107] Where I represents the grayscale value of the radar target 2D image, and O represents the grayscale value of the SAR grayscale image. min_count I represents the grayscale value that appears least frequently in a two-dimensional radar target image. max_count This represents the grayscale value that appears most frequently in a two-dimensional radar target image.

[0108] In one embodiment, the training set is subjected to random data augmentation and normalization to obtain a normalized training set, including:

[0109] The training set is subjected to random data augmentation and normalization based on image rotation, Gaussian white noise perturbation, random uniform noise replacement, and a pre-set random data augmentation method implementation rate, resulting in a normalized training set; the implementation rates of each method in the random data augmentation are 0.3:0.2:0.2, respectively.

[0110] In one embodiment, a deep learning model is built based on the PyTorch framework. Table 1 shows the training and test set settings for three different scenarios in the MSTAR dataset. The actual test data consists of radar targets from three different scenarios in the MSTAR dataset. The recognition rate and variance test algorithm performance were calculated using ten repeated experiments. Three methods were selected for comparison: A-ConvNet, MVGGNet, and Extended Convolutional Capsule Network (ECCNet).

[0111] Table 1

[0112]

[0113]

[0114] To verify the generalization performance of the detection task, a multi-source, multi-scale SAR ship slice dataset was used for verification. The SAR images in the dataset were sourced from the domestic Gaofen-3 satellite and the European Space Agency's Sentinel-1 satellite.

[0115] The network performance was tested using 128×128 SAR image tiles from three different scenarios on the MSTAR dataset. The training set used scene 1 data, and the test set used scenes 2 and 3 data. Table 2 shows the overall recognition rate (OA) and standard deviation (STD) of other methods and the recognition results of this invention on the test set. The experimental data demonstrates that this invention can effectively extract the spatial information of the target in the scene, remove the background clutter effects of different scenes, and still maintain good recognition performance even when the training and test sets are in different scenarios.

[0116] Table 2

[0117]

[0118] The detection results of the MSTAR dataset are as follows: Figure 4 As shown in the figure, the top row is the SAR grayscale image from the MSTAR dataset, and the bottom row is the corresponding detection result of this invention. The detection module of this invention can detect the target region from different background clutter in different scenes, and can effectively separate the target region from the background clutter.

[0119] To verify the generalization performance of the detection module of this invention, a model trained on the MSTAR dataset was used to detect SAR images of ships on the sea surface, verifying the generalization performance of this invention for unknown environments and unknown targets. The multi-source, multi-scale SAR ship slice dataset is 256×256 in size, and the detection results are as follows. Figure 5 As shown, the top row is the SAR grayscale image of the multi-source multi-scale SAR ship slice dataset, and the bottom row is the corresponding detection result of the present invention. The present invention can detect multiple targets from sea clutter of different intensities, and it can be found that it ignores large targets near the shore and only detects the near shore edge.

[0120] Experimental results show that this application utilizes the saliency map of a pre-trained model as a pseudo-label for saliency detection to transfer key feature region information of the target, and uses a feature layer target region mask to separate the target from the background, enabling the model to effectively extract the spatial information of the target in the scene. It can be seen that the method of this invention has good recognition performance in different scenes and can detect the location of targets in unknown scenes.

[0121] 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 in this document, 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.

[0122] In one embodiment, such as Figure 6 As shown, a radar target SAR image joint detection and recognition device based on saliency migration is provided, including: a preprocessing module 602, a random enhancement and normalization module 604, a model building and training module 606, and a recognition module 608, wherein:

[0123] The preprocessing module 602 is used to acquire a two-dimensional image of the radar target to be identified; and to preprocess the two-dimensional image of the radar target to obtain a SAR grayscale image.

[0124] The random augmentation and normalization module 604 is used to divide the SAR grayscale image into a training set and a test set, and to perform random data augmentation and normalization on the training set to obtain the normalized training set.

[0125] The model building and training module 606 is used to build a deep learning model; the deep learning model includes a two-layer convolution, capsule network, saliency detection and feature layer mask; the deep learning model has a U-shaped structure; the deep learning model is trained using a normalized training set to obtain a trained deep learning model;

[0126] The recognition module 608 is used to recognize the test set based on the trained deep learning model and obtain the recognition results.

[0127] In one embodiment, the model building and training module 606 is further configured to train the deep learning model using the normalized training set to obtain a trained deep learning model, including:

[0128] Multiple different training set samples are obtained from the normalized training set; target information is extracted from the different training set samples using double convolution to obtain multi-scale target features;

[0129] Based on the U-shaped structure of the deep learning model, feature aggregation is performed on multi-scale target features to obtain enhanced multi-scale target features;

[0130] The target importance information is transferred from the optical pre-trained model by using the saliency map as a pseudo-label for the saliency detection to obtain the target key region;

[0131] Based on the feature layer mask, target and background separation is performed on the enhanced multi-scale target features to obtain the target features;

[0132] The target features are input into the capsule network for feature transformation to obtain the target spatial features;

[0133] The deep learning model is trained by using the target key region, target spatial features and a pre-set loss function to obtain a trained deep learning model.

[0134] In one embodiment, the process of obtaining a saliency map includes:

[0135] The VGG16 model pre-trained on ImageNet was trained on the standard working condition dataset of MSTAR to obtain a trained VGG16 model.

[0136] Various saliency maps are extracted from the trained VGG16 model using Guided Grad-Cam.

[0137] In one embodiment, the pre-defined loss function includes the capsule network's margin loss, the saliency detection binary classification cross-entropy loss, and the target region mask L1 regularization.

[0138] In one embodiment, the pre-set loss function is:

[0139] L = L m +α·L S +β·L1

[0140] L j =T j max(0,m + -||v j ||) 2 +

[0141] λ(1-T j max(0,||v) j ||-m - ) 2

[0142]

[0143] Where, m + =0.9 represents the lower bound for true positives, m - =0.1 represents the upper bound of the true negative, λ=0.5 represents the proportionality coefficient, L m L represents the interval loss of the capsule network. S L1 represents the cross-entropy loss for saliency detection binary classification, and L1 represents the 1-norm regularization of the target region mask. j Let T represent the loss of class j. j T represents the indicator function. j=1 is only for the j-th class, with hyperparameters α set to 1e-5 and β set to 5e-6 to control the sparsity of the mask.

[0144] In one embodiment, the preprocessing module 602 is further configured to preprocess the two-dimensional image of the radar target to obtain a SAR grayscale image, including:

[0145] A linear transformation and grayscale enhancement are performed on the two-dimensional image of the radar target to obtain a SAR grayscale image;

[0146] The calculation process of linear transformation is as follows:

[0147]

[0148] Where x represents the original radar target 2D image pixel input value, and y represents the output image pixel value, x max x represents the maximum pixel value of the original two-dimensional radar target image. min This represents the minimum pixel value in the original two-dimensional radar target image;

[0149] The calculation process for grayscale enhancement is as follows:

[0150]

[0151] Where I represents the grayscale value of the radar target 2D image, and O represents the grayscale value of the SAR grayscale image. min_count I represents the grayscale value that appears least frequently in a two-dimensional radar target image. max_count This represents the grayscale value that appears most frequently in a two-dimensional radar target image.

[0152] In one embodiment, the random augmentation and normalization module 604 is further configured to perform random data augmentation and normalization processing on the training set to obtain a normalized training set, including:

[0153] The training set is subjected to random data augmentation and normalization based on image rotation, Gaussian white noise perturbation, random uniform noise replacement, and a pre-set random data augmentation method implementation rate, resulting in a normalized training set; the implementation rates of each method in the random data augmentation are 0.3:0.2:0.2, respectively.

[0154] Specific limitations regarding the radar target SAR image joint detection and recognition device based on saliency migration can be found in the limitations of the radar target SAR image joint detection and recognition method based on saliency migration described above, and will not be repeated here. Each module in the aforementioned radar target SAR image joint detection and recognition device based on saliency migration 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.

[0155] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices 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 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 for communication with external terminals via a network connection. When executed by the processor, the computer program implements a joint detection and recognition method for radar target SAR images based on saliency migration. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0156] Those skilled in the art will understand that Figure 7 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.

[0157] 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.

[0158] In one embodiment, a computer storage medium is provided that stores a computer program, which, when executed by a processor, implements the steps of the method described above.

[0159] 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.

[0160] 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.

[0161] 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 joint detection and recognition method for radar target SAR images based on saliency migration, characterized in that, The method includes: Acquire a two-dimensional image of the radar target to be identified; The two-dimensional image of the radar target is preprocessed to obtain a SAR grayscale image; The SAR grayscale image is divided into a training set and a test set. The training set is subjected to random data augmentation and normalization to obtain a normalized training set. Construct a deep learning model; the deep learning model includes two-layer convolution, capsule network, saliency detection, and feature layer masking; the deep learning model has a U-shaped structure; The deep learning model is trained using the normalized training set to obtain a trained deep learning model. The trained deep learning model is used to identify the test set, and the identification results are obtained. The deep learning model is trained using the normalized training set to obtain a trained deep learning model, including: Multiple different training set samples are obtained from the normalized training set; The target information is extracted from the different training set samples using the double-layer convolution to obtain multi-scale target features; Based on the U-shaped structure of the deep learning model, feature aggregation is performed on multi-scale target features to obtain enhanced multi-scale target features; The target importance information is transferred from the optical pre-trained model by using the saliency map as a pseudo-label for the saliency detection to obtain the target key region; Based on the feature layer mask, the enhanced multi-scale target features are separated into target and background to obtain target features; The target features are input into the capsule network for feature transformation to obtain the target spatial features; The deep learning model is trained using the target key region, target spatial features, and a pre-set loss function to obtain a trained deep learning model.

2. The method according to claim 1, characterized in that, The process of obtaining a saliency map includes: The VGG16 model pre-trained on ImageNet was trained on the standard working condition dataset of MSTAR to obtain a trained VGG16 model. Various saliency maps are extracted from the trained VGG16 model using Guided Grad-Cam.

3. The method according to any one of claims 1 to 2, characterized in that, The pre-set loss function includes the capsule network's margin loss, the saliency detection binary classification cross-entropy loss, and the target region mask. Regularization.

4. The method according to claim 1, characterized in that, The pre-set loss function is: in, Indicates the lower bound of a true positive. Indicates the upper bound of the true negative. Represents the proportionality coefficient. This represents the interval loss of the capsule network. This represents the cross-entropy loss for saliency detection in binary classification. This represents the 1-norm regularization of the target region mask. Indicates the first Class loss, Indicates an indicator function, Only when it is the first Class, hyperparameters Set to 1e-5, Set to 5e-6 to control the sparsity of the mask.

5. The method according to claim 1, characterized in that, The radar target two-dimensional image is preprocessed to obtain a SAR grayscale image, including: A linear transformation and grayscale enhancement are performed on the two-dimensional image of the radar target to obtain a SAR grayscale image; The calculation process for the linear transformation is as follows: in, This represents the pixel input values ​​of the original two-dimensional radar target image. Indicates the pixel values ​​of the output image. This represents the maximum pixel value of the original two-dimensional radar target image. This represents the minimum pixel value in the original two-dimensional radar target image; The calculation process for grayscale enhancement is as follows: in, Represents the grayscale value of a two-dimensional radar target image. Represents the grayscale value of a SAR grayscale image. This represents the grayscale value that appears least frequently in a two-dimensional radar target image. This represents the grayscale value that appears most frequently in a two-dimensional radar target image.

6. The method according to claim 1, characterized in that, The training set is subjected to random data augmentation and normalization to obtain a normalized training set, including: The training set is subjected to random data augmentation and normalization processing based on image rotation, Gaussian white noise perturbation, random uniform noise replacement, and a pre-set random data augmentation method implementation rate to obtain a normalized training set; the implementation rates of each method in the random data augmentation are 0.3:0.2:0.2, respectively.

7. A radar target SAR image joint detection and recognition device based on saliency migration, characterized in that, The device includes: The preprocessing module is used to acquire a two-dimensional image of the radar target to be identified; and to preprocess the two-dimensional image of the radar target to obtain a SAR grayscale image. The random augmentation and normalization module is used to divide the SAR grayscale image into a training set and a test set, and to perform random data augmentation and normalization processing on the training set to obtain a normalized training set. A model building and training module is used to build a deep learning model; the deep learning model includes a two-layer convolution, a capsule network, saliency detection, and a feature layer mask; the deep learning model is trained using the normalized training set to obtain a trained deep learning model, including: Multiple different training set samples are obtained from the normalized training set; The target information is extracted from the different training set samples using the double-layer convolution to obtain multi-scale target features; Based on the U-shaped structure of the deep learning model, feature aggregation is performed on multi-scale target features to obtain enhanced multi-scale target features; The target importance information is transferred from the optical pre-trained model by using the saliency map as a pseudo-label for the saliency detection to obtain the target key region; Based on the feature layer mask, the enhanced multi-scale target features are separated into target and background to obtain target features; The target features are input into the capsule network for feature transformation to obtain the target spatial features; The deep learning model is trained using the target key region, target spatial features, and a pre-set loss function to obtain a trained deep learning model; The recognition module is used to recognize the test set according to the trained deep learning model and obtain the recognition result.

8. 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 6.

9. 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 steps of the method according to any one of claims 1 to 6.

Images