Robust Recognition Method and Device for Ground Target SAR Images Based on Contrastive Learning

Abstract

The present application relates to a robust recognition method and device for ground target SAR images based on contrast learning. The method includes: performing gray-scale enhancement on the two-dimensional radar target image to obtain a SAR gray-scale image; dividing the SAR gray-scale image into a training set and a test set, performing random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample; constructing a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrast learning module; training the robust recognition model with different positive example samples to obtain a trained robust recognition model; and performing recognition on the test set according to the trained robust recognition model to obtain a recognition result. Using this method can improve the recognition accuracy of radar ground targets.

Description

Technical Field

[0001] This application relates to the technical field of radar target recognition, and in particular to a robust recognition method, device, computer device and storage medium for ground target SAR images based on contrast learning. Background Art

[0002] Synthetic Aperture Radar (SAR) has the imaging advantages of all-weather and all-day, and is widely used in the fields of automatic target recognition, situation awareness, earth remote sensing, etc.

[0003] Currently, deep learning has been widely used in the field of automatic radar ground target recognition, and has achieved good performance when the imaging conditions of the training set and the test set are less different. However, in practical applications, it faces the problem of non-cooperative target recognition. Due to the non-cooperativeness of the target, the signal-to-noise ratio of its test set is usually inconsistent with that of the training set, and the training sample size is often small. These two problems will affect the recognition performance of the deep learning model, and the recognition accuracy of radar ground targets is low. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a robust recognition method, device, computer device and storage medium for ground target SAR images based on contrast learning, which can improve the recognition accuracy of radar ground targets.

[0005] A robust recognition method for ground target SAR images based on contrast learning, the method includes:

[0006] Obtain a two-dimensional image of the radar target to be recognized;

[0007] Perform gray-scale enhancement on the two-dimensional image of the radar target to obtain a SAR gray-scale image;

[0008] Divide the SAR gray-scale image into a training set and a test set, perform random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample;

[0009] Construct a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network and a contrast learning module;

[0010] Train the robust recognition model with different positive example samples to obtain a trained robust recognition model;

[0011] According to the trained robust recognition model, recognize the test set to obtain a recognition result.

[0012] In one embodiment, the capsule network includes an initial capsule layer and a routing capsule layer; training the robust recognition model with different positive example samples to obtain a trained robust recognition model includes:

[0013] Extract target information from different positive example samples using double-layer convolution to obtain multi-scale target features;

[0014] Input the multi-scale target features into a capsule network for feature transformation to obtain spatial features; the spatial features are positive example features;

[0015] Calculate the positive example features of the spatial features according to the projection layer and prediction layer of the contrast learning module to obtain asymmetric positive example features; construct a contrast learning loss function using the asymmetric positive example features;

[0016] Train a robust recognition model using the contrast learning loss function and the loss function of the capsule network to obtain a trained robust recognition model.

[0017] In one embodiment, inputting the multi-scale target features into a capsule network for feature transformation to obtain spatial features includes:

[0018] Inputting the multi-scale target features into a capsule network for feature transformation, and the obtained spatial features are

[0019]

[0020]

[0021]

[0022] Among them, u i represents the output vector of the initial capsule layer, W ij represents the weight matrix, represents the prediction vector of the routing capsule layer, c ij represents the coupling coefficient, whose value is determined by the dynamic routing algorithm, s j represents the weighted sum of the prediction vectors of the routing capsule layer, v j is the output classification vector of the routing capsule layer, and the output classification vector is the spatial feature.

[0023] In one embodiment, the loss function of the capsule network is

[0024] L j = T j max(0, m + - ||v j ||) 2 +

[0025] λ(1 - T j )max(0, ||v j || - m - ) 2

[0026]

[0027] Among them, m + = 0.9 represents the lower bound of true positive, m - = 0.1 represents the upper bound of true negative, λ = 0.5 represents the proportionality coefficient, and L m represents the loss function of the capsule network, and L j represents the loss of the j-th class, and T j represents the indicator function, and T j = 1 only when it is the j-th class.

[0028] In one embodiment, the contrast learning loss function is

[0029]

[0030]

[0031] Among them, z represents the output feature of the projection layer, h represents the output feature of the prediction layer, and D(h1, z2) represents the similarity between the positive example feature h1 and the positive example feature z2.

[0032] In one embodiment, gray-scale enhancement is performed on the two-dimensional radar target image to obtain a SAR gray-scale image, including:

[0033] Performing gray-scale enhancement on the two-dimensional radar target image, the obtained SAR gray-scale image is

[0034]

[0035] Among them, I represents the gray-scale value of the two-dimensional radar target image, O represents the gray-scale value of the SAR gray-scale image, and I min_count represents the gray-scale value with the least occurrence times in the two-dimensional radar target image, and I max_count represents the gray-scale value with the most occurrence times in the two-dimensional radar target image.

[0036] In one embodiment, random data augmentation and normalization processing are performed on the training set to obtain different positive example samples of the same sample, including:

[0037] Performing random data augmentation and normalization processing on the training set according to image rotation, Gaussian white noise perturbation, random uniform noise replacement, and the implementation rate of the preset random data augmentation method to obtain different positive example samples of the same sample; the implementation rates of each method in the random data augmentation are 0.3:0.2:0.2 respectively.

[0038] A robust recognition device for ground target SAR images based on contrast learning, the device includes:

[0039] A preprocessing module, configured to obtain a two-dimensional radar target image to be recognized; perform gray-scale enhancement on the two-dimensional radar target image to obtain a SAR gray-scale image;

[0040] A data augmentation module, which is used to divide the SAR grayscale image into a training set and a test set, perform random data augmentation and normalization on the training set, and obtain different positive example samples of the same sample;

[0041] A model construction and training module, which is used to construct a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrast learning module; use different positive example samples to train the robust recognition model to obtain a trained robust recognition model;

[0042] An image recognition module, which is used to recognize the test set according to the trained robust recognition model to obtain a recognition result.

[0043] A computer device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the following steps are implemented:

[0044] Obtain a two-dimensional radar target image to be recognized;

[0045] Perform grayscale enhancement on the two-dimensional radar target image to obtain a SAR grayscale image;

[0046] Divide the SAR grayscale image into a training set and a test set, perform random data augmentation and normalization on the training set, and obtain different positive example samples of the same sample;

[0047] Construct a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrast learning module;

[0048] Use different positive example samples to train the robust recognition model to obtain a trained robust recognition model;

[0049] Recognize the test set according to the trained robust recognition model to obtain a recognition result.

[0050] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0051] Obtain a two-dimensional radar target image to be recognized;

[0052] Perform grayscale enhancement on the two-dimensional radar target image to obtain a SAR grayscale image;

[0053] Divide the SAR grayscale image into a training set and a test set, perform random data augmentation and normalization on the training set, and obtain different positive example samples of the same sample;

[0054] Construct a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrast learning module;

[0055] The robust recognition model is trained using different positive example samples to obtain a trained robust recognition model;

[0056] The test set is recognized according to the trained robust recognition model to obtain recognition results.

[0057] The above-mentioned robust recognition method, device, computer equipment and storage medium for ground target SAR images based on contrast learning first divide the SAR grayscale image into a training set and a test set, perform random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample, and construct a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network and a contrast learning module; the robust recognition model is trained using different positive example samples to obtain a trained robust recognition model; the test set is recognized according to the trained robust recognition model to obtain recognition results. By repeatedly using the double-layer convolution, the network depth is increased, and multi-scale target information is provided. The feature vectors extracted by the capsule network have spatial information, and contrast learning further narrows the similarity of different positive examples in the feature space and reduces the intra-class distance. It still has good recognition performance in the case of few samples, and contrast learning is used to increase the spatial similarity of different positive example features, and has a certain invariance to the perturbation form of random data augmentation, thereby improving the robustness against the SNR mismatch problem. Data augmentation represents the prior information of the perturbation form between the training set and the test set, which can be adjusted according to the actual situation, and contrast learning can further improve the data augmentation effect, making the adjustable range of data augmentation parameters increase and improving the actual application effect. Description of the Drawings

[0058] Figure 1 It is a schematic flowchart of a robust recognition method for ground target SAR images based on contrast learning in one embodiment;

[0059] Figure 2 It is an example diagram of the grayscale image of the ground target SAR image in one embodiment;

[0060] Figure 3 It is a structure diagram of the robust recognition model in one embodiment;

[0061] Figure 4 It is a structure diagram of the double-layer convolution in another embodiment;

[0062] Figure 5 It is a structure diagram of the initial capsule layer in one embodiment;

[0063] Figure 6 It is a structure diagram of the contrast learning module in one embodiment;

[0064] Figure 7The structural diagram of the projection layer in the contrastive learning module in one embodiment;

[0065] Figure 8 The structural diagram of the prediction layer in the contrastive learning module in one embodiment;

[0066] Figure 9 The structural block diagram of a ground target SAR image robust recognition device based on contrastive learning in one embodiment;

[0067] Figure 10 The internal structural diagram of a computer device in one embodiment. Detailed implementation manners

[0068] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0069] In one embodiment, as Figure 1 shown, a ground target SAR image robust recognition method based on contrastive learning is provided, including the following steps:

[0070] Step 102, obtain a two-dimensional radar target image to be recognized; perform gray-scale enhancement on the two-dimensional radar target image to obtain a SAR gray-scale image.

[0071] Step 104, divide the SAR gray-scale image into a training set and a test set, perform random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample.

[0072] Performing random data augmentation on the training set to obtain different positive examples of the same sample, and performing normalization as training samples. The random data augmentation methods include image rotation, Gaussian white noise perturbation, and random uniform noise replacement. The parameters and random probabilities of each method can be adjusted. Data augmentation represents the prior information of the perturbation form between the training set and the test set, which can be adjusted according to the actual situation. Subsequently, contrastive learning can be used to further improve the data augmentation effect, increase the adjustable range of data augmentation parameters, and improve the actual application effect.

[0073] Step 106, construct a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrastive learning module; use different positive example samples to train the robust recognition model to obtain a trained robust recognition model.

[0074] Constructing the robust recognition model is as Figure 3 shown, and the double-layer convolution is as Figure 4As shown, by repeatedly using double-layer convolution, the network depth is increased, and multi-scale target features are provided. The capsule network includes an initial capsule layer and a routing capsule layer. The structure of the initial capsule layer is as Figure 5 shown. The routing capsule layer is similar to a fully connected neural network and is represented in vector form in this application. The capsule network extracts the spatial information of the target features. The structure of the contrast learning module is as Figure 6 shown. The vector features extracted by the capsule network are input into the projection layer and the prediction layer to calculate the spatial similarity of different positive example features. The projection layer is as Figure 7 shown, and the structure of the prediction layer is as Figure 8 shown. The projection layer increases the network depth and extracts more general feature information. The prediction layer provides an asymmetric structure and stop gradient to avoid using negative examples together, reducing the computational amount. Through contrast learning, the spatial similarity of different positive example features is increased, the intra-class distance is reduced, and the robustness of the model is improved. High recognition rates of radar ground targets can be achieved under a wide range of signal-to-noise ratio conditions, and good performance can still be obtained when the training sample size is small.

[0075] Step 108, identify the test set according to the trained robust recognition model to obtain the recognition result.

[0076] In the above-mentioned robust recognition method for ground target SAR images based on contrast learning, first, 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 processing to obtain different positive example samples of the same sample, and a robust recognition model is constructed. The robust recognition model includes double-layer convolution, a capsule network, and a contrast learning module. The robust recognition model is trained using different positive example samples to obtain a trained robust recognition model. The test set is identified according to the trained robust recognition model to obtain the recognition result. By repeatedly using double-layer convolution, the network depth is increased, and multi-scale target information is provided. The feature vectors extracted by the capsule network have spatial information. Contrast learning further narrows the similarity of different positive examples in the feature space and reduces the intra-class distance. Good recognition performance can still be obtained when the number of samples is small. And contrast learning is used to increase the spatial similarity of different positive example features, which has a certain invariance to the perturbation form of random data augmentation, thereby improving the robustness against the signal-to-noise ratio mismatch problem. Data augmentation represents the prior information of the perturbation form between the training set and the test set, which can be adjusted according to the actual situation. And contrast learning can further improve the data augmentation effect, increasing the adjustable range of data augmentation parameters and improving the actual application effect.

[0077] In one of the embodiments, the capsule network includes an initial capsule layer and a routing capsule layer; the robust recognition model is trained using different positive example samples to obtain a trained robust recognition model, including:

[0078] Use double-layer convolution to extract target information from different positive example samples to obtain multi-scale target features;

[0079] Input the multi-scale target features into the capsule network for feature transformation to obtain spatial features; the spatial features are positive example features;

[0080] Calculate the positive example features of the spatial features according to the projection layer and prediction layer of the contrast learning module to obtain asymmetric positive example features; use the asymmetric positive example features to construct a contrast learning loss function;

[0081] Use the contrast learning loss function and the loss function of the capsule network to train the robust recognition model to obtain a trained robust recognition model.

[0082] In a specific embodiment, use the asymmetric positive example features to construct a contrast learning loss function to increase the spatial similarity of different positive example features; the network training parameters are as follows: the training batch is 100, the batch training size is 64, the learning rate is 3e-4, the exponential decay rate is 0.98, and the optimizer uses the NAdam algorithm.

[0083] In one embodiment, input the multi-scale target features into the capsule network for feature transformation to obtain spatial features, including:

[0084] Input the multi-scale target features into the capsule network for feature transformation, and the obtained spatial features are

[0085]

[0086]

[0087]

[0088] Among them, u i represents the output vector of the initial capsule layer, W ij represents the weight matrix, represents the prediction vector of the routing capsule layer, c ij represents the coupling coefficient, and its value is determined by the dynamic routing algorithm, s j represents the weighted sum of the prediction vectors of the routing capsule layer, v j is the output classification vector of the routing capsule layer, and the output classification vector is the spatial feature. The angle of the vector has spatial information, and the vector product between two vectors represents the spatial similarity.

[0089] In a specific embodiment, according to and v j Adjust c according to the spatial similarity, that is, the vector product ij to extract the spatial information of the target features.

[0090] In one embodiment, the loss function of the capsule network is

[0091] L j = T j max(0, m + - ||v j ||) 2 +

[0092] λ(1 - T j )max(0, ||v j || - m - ) 2

[0093]

[0094] where m + = 0.9 represents the lower bound of true positives, m - = 0.1 represents the upper bound of true negatives, λ = 0.5 represents the proportionality coefficient, L m represents the loss function of the capsule network, L j represents the loss of the j-th class, T j represents the indicator function, T j = 1 only when it is the j-th class.

[0095] In a specific embodiment, the loss function of the capsule network can be applied to the multi-classification situation, expanding the application scenario of the present invention. And by restricting the ranges of true positives and true negatives, the probabilities of false positives and false negatives are reduced, improving the recognition performance.

[0096] In one embodiment, the contrastive learning loss function is

[0097]

[0098]

[0099] where z represents the output feature of the projection layer, h represents the output feature of the prediction layer, D(h1, z2) represents the similarity between the positive example feature h1 and the positive example feature z2, and the positive example feature h1 and the positive example feature z2 are asymmetric positive example features.

[0100] In a specific embodiment, the contrastive learning loss function constrains the perturbation range of data augmentation for features by increasing the similarity of features, reduces the intra-class distance of different positive examples, enhances the stability of spatial feature extraction of the capsule network under different perturbation situations, and enhances the recognition robustness.

[0101] In one embodiment, gray-scale enhancement is performed on the two-dimensional radar target image to obtain the SAR gray-scale image, including:

[0102] Perform gray-scale enhancement on the two-dimensional image of the radar target to obtain the SAR gray-scale image as

[0103]

[0104] where I represents the gray-scale value of the two-dimensional image of the radar target, O represents the gray-scale value of the SAR gray-scale image, I min_count represents the gray-scale value with the least number of occurrences in the two-dimensional image of the radar target, I max_count represents the gray-scale value with the most number of occurrences in the two-dimensional image of the radar target.

[0105] In one embodiment, perform random data augmentation and normalization on the training set to obtain different positive example samples of the same sample, including:

[0106] Perform random data augmentation and normalization on the training set according to image rotation, Gaussian white noise perturbation, random uniform noise replacement, and the implementation rate of the preset random data augmentation method to obtain different positive example samples of the same sample; the implementation rates of each method in the random data augmentation are 0.3:0.2:0.2 respectively.

[0107] In a specific embodiment, the implementation probabilities of each method in the random data augmentation are 0.3, 0.2, and 0.2 respectively. The rotation angle of the image rotation in the random data augmentation is between [-5, 5]. The mean of the Gaussian white noise in the random data augmentation is 0.1, the standard deviation is 0.1, and the amplitude is between [0.5, 1.5]. For the case of signal-to-noise ratio mismatch, the mean of the Gaussian white noise is 0, the standard deviation is 0.8, and the amplitude is between [0, 1]. The range of the uniform noise in the random data augmentation is between [0, 1], and the proportion of replacing the original image is between [0, 0.05]. For the case of uniform noise corrosion, the proportion of replacing the original image is between [0, 0.3].

[0108] In one embodiment, the experiment on the measured data further illustrates the beneficial effects of this application;

[0109] Build a deep learning model based on the pytorch framework, and the measured data is the MSTAR dataset, which contains ten types of ground targets, such as Figure 2As shown in the figure. The performance of the verification algorithm was evaluated by calculating the recognition rate and variance through ten repeated experiments. Four types of SAR image recognition methods for ground targets, namely A-ConvNet, AM-CNN, MVGGNet, and Extended convolutional Capsule Network (ECCNet), were selected as the comparison methods. A-ConvNet adopted fully convolutional layers to reduce overfitting. AM-CNN was an improvement on A-ConvNet, adding an attention mechanism and BN layers. MVGGNet used pre-trained model parameters from optical images and considered the target space feature information. Meanwhile, the same data augmentation method was added to the above methods.

[0110] As shown in Table 1, a Standard Operating Condition (SOC) dataset was constructed. The grazing angle of the training set was 17°, and the grazing angle of the test set was 15°. Therefore, the imaging conditions of the two sets were less different.

[0111] Table 1

[0112]

[0113] Under the SOC condition, the network performance was tested. The overall recognition rate (OA) and standard deviation (STD) of the recognition results of other methods and this application are shown in Table 2. Due to the small difference in the datasets, each method performed well, and this application achieved the best performance.

[0114] Table 2

[0115]

[0116] Based on the SOC dataset, experiments were conducted under Extended Operating Condition (EOC) to test the network performance, including signal-to-noise ratio mismatch and uniform noise corrosion. In the case of signal-to-noise ratio mismatch, additive white Gaussian noise with different signal-to-noise ratios was added to the test set for verification. In the case of uniform noise corrosion, different proportions of pixels in the test set were replaced with uniformly distributed noise. The overall recognition rate (OA) and standard deviation (STD) of the recognition results of other methods and this application are shown in Tables 3 and 4. The method proposed in this application can effectively handle signal-to-noise ratio mismatch and uniform noise corrosion, and has obvious robustness compared with other methods under different signal-to-noise ratios and different proportions of corrosion.

[0117] Table 3

[0118]

[0119] Table 4

[0120]

[0121] When reducing the sample size of the training set based on the SOC dataset, the performance of each method is verified under the problem of a small training set sample size. The overall recognition rate (OA) and standard deviation (STD) of the recognition results of other methods and this application are shown in Table 5. This application has good performance under the condition of fewer samples, significantly higher than other methods at 10% of the training sample size, and the recognition rate still reaches more than 90%.

[0122] Table 5

[0123]

[0124] It can be found from the experimental results that this application increases the spatial similarity of different positive examples through contrastive learning, narrows the intra-class distance, and enhances the spatial information of the features extracted by the capsule network; improves the data augmentation effect and has a certain invariance to the data augmentation form. It can be found that a robust recognition method for ground target SAR images based on contrastive learning in this application has good robust recognition performance for the situation of signal-to-noise ratio mismatch of non-cooperative targets and fewer training samples.

[0125] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover,

[0126] In one embodiment, as Figure 9 shown, a robust recognition device for ground target SAR images based on contrastive learning is provided, including: a preprocessing module 902, a data augmentation module 904, a model construction and training module 906, and an image recognition module 908, where:[[]]END]]

[0127] The preprocessing module 902 is used to obtain a two-dimensional radar target image to be recognized; perform gray-scale enhancement on the two-dimensional radar target image to obtain a SAR gray-scale image;

[0128] The data augmentation module 904 is used to divide the SAR gray-scale image into a training set and a test set, perform random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample;

[0129] The model construction and training module 906 is used to construct a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrast learning module; different positive example samples are used to train the robust recognition model to obtain a trained robust recognition model;

[0130] The image recognition module 908 is used to recognize the test set according to the trained robust recognition model to obtain a recognition result.

[0131] In one embodiment, the capsule network includes an initial capsule layer and a routing capsule layer; the model construction and training module 906 is further used to train the robust recognition model with different positive example samples to obtain a trained robust recognition model, including:

[0132] Using the double-layer convolution to extract target information from different positive example samples to obtain multi-scale target features;

[0133] Inputting the multi-scale target features into the capsule network for feature transformation to obtain spatial features; the spatial features are positive example features;

[0134] Calculating positive example features for the spatial features according to the projection layer and the prediction layer of the contrast learning module to obtain asymmetric positive example features; constructing a contrast learning loss function using the asymmetric positive example features;

[0135] Training the robust recognition model using the contrast learning loss function and the loss function of the capsule network to obtain a trained robust recognition model.

[0136] In one embodiment, the model construction and training module 906 is further used to input the multi-scale target features into the capsule network for feature transformation to obtain spatial features, including:

[0137] Inputting the multi-scale target features into the capsule network for feature transformation, and the obtained spatial features are

[0138]

[0139]

[0140]

[0141] where, u i represents the output vector of the initial capsule layer, W ij represents the weight matrix, represents the prediction vector of the routing capsule layer, c ij represents the coupling coefficient, whose value is determined by the dynamic routing algorithm, s j represents the weighted sum of the prediction vectors of the routing capsule layer, v j is the output classification vector of the routing capsule layer, and the output classification vector is the spatial feature.

[0142] In one embodiment, the loss function of the capsule network is

[0143] L j = T j max(0, m + - ||v j ||) 2 +

[0144] λ(1 - T j )max(0, ||v j || - m - ) 2

[0145]

[0146] where m + = 0.9 represents the lower bound of true positives, m - = 0.1 represents the upper bound of true negatives, λ = 0.5 represents the proportionality coefficient, L m represents the loss function of the capsule network, L j represents the loss of the j-th class, T j represents the indicator function, T j = 1 only when it is the j-th class.

[0147] In one embodiment, the contrastive learning loss function is

[0148]

[0149]

[0150] where z represents the output feature of the projection layer, h represents the output feature of the prediction layer, and D(h1, z2) represents the similarity between the positive example feature h1 and the positive example feature z2.

[0151] In one embodiment, the data augmentation module 904 is further configured to perform grayscale enhancement on the two-dimensional radar target image to obtain a SAR grayscale image, including:

[0152] Performing grayscale enhancement on the two-dimensional radar target image, the obtained SAR grayscale image is

[0153]

[0154] where I represents the grayscale value of the two-dimensional radar target image, O represents the grayscale value of the SAR grayscale image, I min_count represents the grayscale value with the least occurrence times in the two-dimensional radar target image, I max_count represents the grayscale value with the most occurrence times in the two-dimensional radar target image.

[0155] In one embodiment, the data augmentation module 904 is further configured to perform random data augmentation and normalization on the training set to obtain different positive example samples of the same sample, including:

[0156] Performing random data augmentation and normalization on the training set according to image rotation, Gaussian white noise perturbation, random uniform noise replacement, and a preset implementation rate of the random data augmentation method to obtain different positive example samples of the same sample; the implementation rates of the respective methods in the random data augmentation are 0.3:0.2:0.2.

[0157] For the specific limitations of a ground target SAR image robust recognition device based on contrast learning, reference can be made to the limitations of a ground target SAR image robust recognition method based on contrast learning in the above text, which will not be elaborated here. Each module in the above-mentioned ground target SAR image robust recognition device based on contrast learning can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the above-mentioned modules.

[0158] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as Figure 10 shown. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a ground target SAR image robust recognition method based on contrast learning. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device may be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0159] Those skilled in the art can understand that Figure 10 the structure shown in

[0160] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method in the above embodiment are implemented.

[0161] In one embodiment, a computer storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in the above embodiment are implemented.

[0162] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories 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), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0163] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, 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, it should be considered to be within the scope described in this specification.

Claims

1. A robust recognition method for ground target SAR images based on contrastive learning, characterized in that, The method includes: Obtaining a two-dimensional image of a radar target to be recognized; Performing gray-scale enhancement on the two-dimensional image of the radar target to obtain a SAR gray-scale image; Dividing the SAR gray-scale image into a training set and a test set, and performing random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample; Constructing a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrast learning module; Training the robust recognition model with the different positive example samples to obtain a trained robust recognition model; Recognizing the test set according to the trained robust recognition model to obtain a recognition result; The capsule network includes an initial capsule layer and a routing capsule layer; training the robust recognition model with the different positive example samples to obtain a trained robust recognition model, including: Extracting target information from the different positive example samples using the double-layer convolution to obtain multi-scale target features; Inputting the multi-scale target features into the capsule network for feature transformation to obtain spatial features; the spatial features are positive example features; Calculating positive example features for the spatial features according to the projection layer and the prediction layer of the contrast learning module to obtain asymmetric positive example features; constructing a contrast learning loss function using the asymmetric positive example features; Training the robust recognition model using the contrast learning loss function and the loss function of the capsule network to obtain a trained robust recognition model.

2. According to the method described in claim 1, inputting the multi-scale target features into the capsule network for feature transformation to obtain spatial features, including: Inputting the multi-scale target features into the capsule network for feature transformation, and the obtained spatial features are Among them, u i represents the output vector of the initial capsule layer, W ij represents the weight matrix, represents the predicted vector of the routing capsule layer, c ij represents the coupling coefficient, whose value is determined by the dynamic routing algorithm, s j represents the weighted sum of the predicted vectors of the routing capsule layer, v j is the output classification vector of the routing capsule layer, and the output classification vector is a spatial feature.

3. According to the method described in claim 2, the loss function of the capsule network is L j = T j max(0, m + - ||v j ||) 2 + λ(1 - T j ) max(0, ||v j || - m - ) 2 Among them, m + = 0.9 represents the lower bound of true positives, m - = 0.1 represents the upper bound of true negatives, λ = 0.5 represents the proportionality coefficient, L m represents the loss function of the capsule network, L j represents the loss of the j-th class, T j represents the indicator function, T j = 1 only when it is the j-th class.

4. According to the method described in claim 1, the contrast learning loss function is Among them, z represents the output feature of the projection layer, h represents the output feature of the prediction layer, and D(h1,z2) represents the similarity between the positive example feature h1 and the positive example feature z2.

5. The method according to claim 1, wherein Performing gray-scale enhancement on the two-dimensional image of the radar target to obtain a SAR gray-scale image, including: Performing gray-scale enhancement on the two-dimensional image of the radar target, and the obtained SAR gray-scale image is Among them, I represents the gray value of the two-dimensional image of the radar target, O represents the gray value of the SAR gray image, and I min_count represents the gray value with the least number of occurrences in the two-dimensional image of the radar target, and I max_count represents the gray value with the most number of occurrences in the two-dimensional image of the radar target.

6. The method according to claim 1, wherein Performing random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample, including: Performing random data augmentation and normalization processing on the training set according to image rotation, Gaussian white noise perturbation, random uniform noise replacement, and a pre-set random data augmentation method implementation rate to obtain different positive example samples of the same sample; the implementation rates of the respective methods in the random data augmentation are 0.3:0.2:0.

2.

7. A robust recognition device for ground target SAR images based on contrastive learning, characterized in that, The device includes: A preprocessing module, configured to obtain a two-dimensional image of a radar target to be recognized; perform gray-scale enhancement on the two-dimensional image of the radar target to obtain a SAR gray-scale image; A data augmentation module, configured to divide the SAR gray-scale image into a training set and a test set, and perform random data augmentation and normalization processing on the training set to obtain different positive example samples of the same sample; A model construction and training module for constructing a robust recognition model; the robust recognition model includes a double-layer convolution, a capsule network, and a contrast learning module; training the robust recognition model with the different positive example samples to obtain a trained robust recognition model; the capsule network includes an initial capsule layer and a routing capsule layer; training the robust recognition model with the different positive example samples to obtain a trained robust recognition model, including: extracting target information from the different positive example samples by using the double-layer convolution to obtain multi-scale target features; inputting the multi-scale target features into the capsule network for feature transformation to obtain spatial features; the spatial features are positive example features; calculating positive example features for the spatial features according to the projection layer and the prediction layer of the contrast learning module to obtain asymmetric positive example features; constructing a contrast learning loss function by using the asymmetric positive example features; training the robust recognition model by using the contrast learning loss function and the loss function of the capsule network to obtain a trained robust recognition model; An image recognition module for recognizing the test set according to the trained robust recognition model to obtain a recognition result.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

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

Images