A SAR Image Target Recognition Method Based on Small Sample Incremental Learning

Abstract

The present invention relates to a SAR image target recognition method based on small sample incremental learning, comprising: step 1: constructing a deep residual network based on the prototype idea; step 2: using a base class data set and an incremental data set to perform base class training and incremental learning on the deep residual network based on the prototype idea respectively; step 3: inputting a SAR image to be tested into the deep residual network based on the prototype idea that has completed training and learning, and obtaining a prediction classification result of the SAR image to be tested. The SAR image target recognition method based on small sample incremental learning of the present invention designs a deep residual network based on the prototype idea, which has the ability to automatically extract SAR image features even when the image samples are small. Compared with the existing technology, the method of mapping the SAR image into the feature space can reduce the overfitting phenomenon caused by the lack of sample quantity to a certain extent, and improve the overall network's recognition ability for small sample images.

Description

Technical Field

[0001] The invention belongs to the field of radar image processing, and in particular relates to a SAR image target recognition method based on small sample incremental learning. Background Art

[0002] Synthetic aperture radar (SAR) is an active microwave imaging radar that has the ability to obtain multi-dimensional surface information all day and all weather, and has a certain ability to penetrate the surface and vegetation. Synthetic aperture radar imaging technology is an indispensable technical means in the current field of earth observation. Automatic target recognition (ATR) is a technology that determines the image category by capturing image feature information. It is still quite difficult to complete real-time automatic target recognition of radar images in a complex electromagnetic environment.

[0003] Currently, deep learning-based methods provide a new approach for automatic target recognition. Deep learning-based methods do not require the establishment of a good model for radar images. Instead, they use images obtained by radar imaging technology to train the model based on the initialization model, allowing the model to learn the characteristics of radar images and complete automatic recognition of radar images.

[0004] Due to the excellent feature extraction and recognition capabilities of deep learning, the application of deep learning methods to SAR ATR has attracted more attention. Many methods have been proposed for automatic target recognition of SAR images. Among them, deep convolutional neural networks have good performance in feature extraction, and feedforward neural networks also have good performance in target recognition.

[0005] Since deep learning training a network often takes a lot of time, and SAR images are difficult to obtain and the cost is not low, when the network encounters a new task or a new class, it is unrealistic to add small samples of the new class to the training samples and then retrain the network. If the network is not retrained, the network cannot recognize the new class and will mistakenly classify it into the known class. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a SAR image target recognition method based on small sample incremental learning. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] The present invention provides a SAR image target recognition method based on small sample incremental learning, comprising:

[0008] Step 1: Construct a deep residual network based on the prototype idea;

[0009] Step 2: Using the base class data set and the incremental data set, the prototype-based deep residual network is respectively subjected to base class training and incremental learning;

[0010] Step 3: Input the SAR image to be tested into the deep residual network based on the prototype idea that has been trained and learned to obtain the prediction and classification result of the SAR image to be tested.

[0011] In one embodiment of the present invention, the deep residual network based on the prototype idea includes a convolution module, a deep residual network, an adaptive average pooling layer, a fully connected layer and a classifier connected in sequence, wherein:

[0012] The convolution module includes a first convolution layer and a first ReLU activation layer connected in sequence;

[0013] The deep residual network includes a plurality of residual blocks connected in sequence, and is used to extract image features of an input image;

[0014] The fully connected layer is used to store class prototypes, and a graph attention network is integrated in the fully connected layer, and the graph attention network is used to adjust the position of the class prototype stored in the fully connected layer in the incremental learning stage;

[0015] The classifier is a cosine classifier, which is used to output a predicted classification result.

[0016] In one embodiment of the present invention, the residual block includes a second convolutional layer, a first batch of normalization layers, a second ReLU activation layer, a third convolutional layer, a second batch of normalization layers and a third ReLU activation layer connected in sequence;

[0017] The input of the residual block and the input of the residual block are sequentially passed through the second convolutional layer, the first batch of normalization layers, the second ReLU activation layer, the third convolutional layer, and the output of the second batch of normalization layers is added and output after passing through the third ReLU activation layer.

[0018] In one embodiment of the present invention, the residual block further includes a channel number conversion unit, the channel number conversion unit is connected between the input end of the second convolutional layer and the output end of the second batch normalization layer, and the channel number conversion unit includes a convolutional layer and a batch normalization layer connected in sequence;

[0019] When the number of input channels of the residual block is not equal to the number of channels of its output, the input of the residual block is converted by the channel number conversion unit to have the same number of channels as its output, and the result of adding the input of the residual block to the output of the second batch normalization layer is output after passing through the third ReLU activation layer.

[0020] In one embodiment of the present invention, the base dataset and the incremental dataset both include a number of SAR images of multiple categories with classification labels, and the incremental dataset is a 1-way-5-shot or 1-way-1-shot small sample dataset.

[0021] In one embodiment of the present invention, step 2 comprises:

[0022] Step 2.1: In the base class training stage, the base class data set is used to train and learn the deep residual network based on the prototype idea, and the network weights of the deep residual network based on the prototype idea are updated through back propagation to obtain the network model of the base class training stage;

[0023] Step 2.2: In the incremental learning stage, the network model of the base class training stage is trained and learned using the incremental data set, and the weights of the fully connected layer are updated through back propagation to obtain a deep residual network based on the prototype idea that has been trained and learned.

[0024] In one embodiment of the present invention, step 2.1 comprises:

[0025] Step 1a: Initialize the network parameters of the deep residual network based on the prototype idea;

[0026] Step 1b: obtaining base class training samples and their corresponding classification labels in the base class data set according to the index, inputting the base class training samples into the deep residual network based on the prototype idea, and obtaining the mapping of the base class training samples in the feature space after passing through the convolution module and the deep residual network;

[0027] Step 1c: the mapping is input into the classifier after passing through the adaptive average pooling layer and the fully connected layer, and the classifier outputs the cosine similarity between the base class training sample and the class prototype;

[0028] Step 1d: taking the class corresponding to the index with the largest cosine similarity as the predicted classification result of the base class training sample, and calculating the cross entropy loss function and accuracy according to the classification label corresponding to the base class training sample;

[0029] Step 1e: According to the cross entropy loss function, the network weights of the deep residual network based on the prototype idea are updated through back propagation, and the mapping expression of the base class training samples in the feature space is taken as the class prototype, and stored as the weight of the fully connected layer to complete the base class training stage of the deep residual network based on the prototype idea and obtain the network model of the base class training stage.

[0030] In one embodiment of the present invention, the step 2.2 comprises:

[0031] Step 2a: Obtain incremental training samples and their corresponding classification labels in the incremental data set, input the incremental training samples into the network model of the base class training stage, and obtain their mapping in the feature space;

[0032] Step 2b: taking the average of the mapping expressions of samples of the same class, obtaining a new class prototype of the new class in the feature space, and using the new class prototype as the newly added weight of the fully connected layer;

[0033] Step 2c: input the base class training samples and the incremental training samples into the network model obtained in step 2b, and calculate the cross entropy loss function according to the base class training samples and the incremental training samples and their corresponding classification labels;

[0034] Step 2d: According to the cross entropy loss function, the weights of the fully connected layer are updated by back propagation to obtain a deep residual network based on the prototype idea that has been trained and learned.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The SAR image target recognition method based on small sample incremental learning of the present invention designs a deep residual network based on the prototype idea. Even when the image samples are small, the network can automatically extract the features of the SAR image. Compared with the existing technology, the method of mapping the SAR image into the feature space can reduce the overfitting phenomenon caused by the lack of sample quantity to a certain extent, and improve the overall network's recognition ability for small sample images;

[0037] 2. The SAR image target recognition method based on small sample incremental learning of the present invention designs a continuously evolving classifier that can continue to learn new classes in the trained network. This setting is more in line with the need to continue to learn small sample new class SAR image information in reality;

[0038] 3. The SAR image target recognition method based on small sample incremental learning of the present invention adopts a graph attention network to store various prototypes, and fine-tunes the distance between these prototypes in the graph attention network to obtain a better classification boundary. This storage method is directly integrated in the fully connected layer of the network and is trained together with the network according to the training samples. It will have a better decision boundary than the network with fixed class prototypes and improve the recognition accuracy.

[0039] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following specifically cites a preferred embodiment and describes it in detail with the accompanying drawings as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic diagram of a SAR image target recognition method based on small sample incremental learning provided by an embodiment of the present invention;

[0041] Figure 2 It is a schematic diagram of the structure of a deep residual network based on the prototype concept provided by an embodiment of the present invention;

[0042] Figure 3 It is a schematic diagram of a process of training and testing a deep residual network based on a prototype concept provided by an embodiment of the present invention;

[0043] Figure 4 It is a structural diagram of the training and learning stages of a graph attention network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, a SAR image target recognition method based on small sample incremental learning proposed by the present invention is described in detail below in combination with the accompanying drawings and specific implementation methods.

[0045] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0046] Embodiment 1

[0047] See also Figure 1 , Figure 1 is a schematic diagram of a SAR image target recognition method based on small sample incremental learning provided by an embodiment of the present invention. As shown in the figure, the SAR image target recognition method based on small sample incremental learning of this embodiment includes:

[0048] Step 1: Construct a deep residual network based on the prototype idea;

[0049] Step 2: Use the base class dataset and incremental dataset to perform base class training and incremental learning on the prototype-based deep residual network respectively;

[0050] Step 3: Input the SAR image to be tested into the deep residual network based on the prototype idea that has been trained and learned to obtain the predicted classification result of the SAR image to be tested.

[0051] Please refer to Figure 2 , Figure 2It is a structural diagram of a deep residual network based on the prototype concept provided by an embodiment of the present invention. As shown in the figure, the deep residual network based on the prototype concept of this embodiment includes a convolution module, a deep residual network, an adaptive average pooling layer, a fully connected layer and a classifier connected in sequence. Among them, the convolution module includes a first convolution layer and a first ReLU activation layer connected in sequence; the deep residual network includes a plurality of residual blocks connected in sequence, which are used to extract image features of the input image; the fully connected layer is used to store class prototypes, and a graph attention network is integrated in the fully connected layer. The graph attention network is used to adjust the position of the class prototype stored in the fully connected layer in the incremental learning stage; the classifier is a cosine classifier, which is used to output the predicted classification result.

[0052] Specifically, the residual block includes a second convolutional layer, a first batch of normalization layers, a second ReLU activation layer, a third convolutional layer, a second batch of normalization layers, and a third ReLU activation layer connected in sequence. The input of the residual block and the output of the residual block after the input passes through the second convolutional layer, the first batch of normalization layers, the second ReLU activation layer, the third convolutional layer, and the second batch of normalization layers in sequence are added and output after passing through the third ReLU activation layer.

[0053] Furthermore, the residual block also includes a channel number conversion unit, which is connected between the input end of the second convolutional layer and the output end of the second batch normalization layer, and the channel number conversion unit includes a convolutional layer and a batch normalization layer connected in sequence. In this embodiment, since the input and output of the residual block are additive, when the number of input channels of the residual block is not equal to the number of channels of its output, the input of the residual block is converted to the same number of channels as its output by the channel number conversion unit, and the result of adding the output of the second batch normalization layer is output after passing through the third ReLU activation layer.

[0054] Furthermore, in this embodiment, both the base dataset and the incremental dataset include a plurality of SAR images of multiple categories with classification labels, and the incremental dataset is a 1-way-5-shot or 1-way-1-shot small sample dataset.

[0055] Optionally, the training samples of the base dataset and the incremental dataset are SAR images obtained from the MSTAR dataset, and the incremental dataset is SAR images observed at different elevation angles.

[0056] It should be noted that since the image data fed into the prototype-based deep residual network at each stage of training and learning are different, it is necessary to obtain the image data index needed at each stage from the MSTAR radar image dataset in advance.

[0057] Further, in this embodiment, step 2 includes:

[0058] Step 2.1: In the base class training phase, the base class data set is used to train the deep residual network based on the prototype idea, and the network weights of the deep residual network based on the prototype idea are updated through back propagation to obtain the network model of the base class training phase;

[0059] Specifically, step 2.1 includes:

[0060] Step 1a: Initialize the network parameters of the deep residual network based on the prototype idea;

[0061] Step 1b: Obtain the base class training samples and their corresponding classification labels in the base class dataset according to the index, input the base class training samples into the deep residual network based on the prototype idea, and obtain the mapping of the base class training samples in the feature space after passing through the convolution module and the deep residual network;

[0062] Step 1c: The mapping is input into the classifier after passing through the adaptive average pooling layer and the fully connected layer, and the classifier outputs the cosine similarity between the base class training sample and the class prototype;

[0063] Step 1d: The class corresponding to the index with the largest cosine similarity is used as the predicted classification result of the base class training sample, and the cross entropy loss function and accuracy are calculated according to the classification label corresponding to the base class training sample;

[0064] Step 1e: According to the cross entropy loss function, the network weights of the deep residual network based on the prototype idea are updated through back propagation, and the mapping expression of the base class training samples in the feature space is taken as the class prototype, which is stored as the weight of the fully connected layer to complete the base class training phase of the deep residual network based on the prototype idea and obtain the network model of the base class training phase.

[0065] Step 2.2: In the incremental learning stage, the network model of the base class training stage is trained and learned using the incremental data set, and the weights of the fully connected layer are updated through back propagation to obtain a deep residual network based on the prototype idea that has been trained and learned.

[0066] Specifically, step 2.2 includes:

[0067] Step 2a: Obtain the incremental training samples and their corresponding classification labels in the incremental data set, input the incremental training samples into the network model of the base class training phase, and obtain their mapping in the feature space;

[0068] Step 2b: Take the average of the mapping expressions of samples of the same class to obtain the new class prototype of the new class in the feature space, and use the new class prototype as the new weight of the fully connected layer;

[0069] Step 2c: Input the base class training samples and the incremental training samples into the network model obtained in step 2b, and calculate the cross entropy loss function according to the base class training samples and the incremental training samples and their corresponding classification labels;

[0070] Step 2d: According to the cross entropy loss function, the weights of the fully connected layer are updated through back propagation to obtain a deep residual network based on the prototype idea that has been trained and learned.

[0071] It should be noted that, in step 3, if the SAR image to be tested involves a new class, incremental learning is required for the SAR image of the new class.

[0072] Furthermore, combined with Figure 3 and Figure 4 , the base class training and incremental learning process of the deep residual network based on the prototype idea of ​​this embodiment is specifically described, Figure 3 is a flowchart of training and testing a deep residual network based on a prototype concept provided by an embodiment of the present invention. Figure 4 It is a structural diagram of the training and learning stages of a graph attention network provided by an embodiment of the present invention.

[0073] In this embodiment, the image data indexes required for each stage are obtained in advance from the MSTAR radar image dataset. The base class dataset is set to 7 categories, and the incremental dataset is set to 3 categories. The incremental training samples in the 3 categories of incremental datasets are divided into three incremental stages for classifier learning. There are enough base class images, and all the training samples of the first 7 categories are used, which is consistent with the general closed set training; the 3 categories of incremental training samples in the incremental learning stage are all trained in a 1-way-5-shot small sample method. At the same time, according to different stages, the test set samples are used as the verification set to ensure the accuracy while minimizing the overfitting problem of small sample learning.

[0074] Specifically, the training set with a pitch angle of 17° in the MSTAR radar image data set is used as the training set of this embodiment, where the first 7 categories are used as base class training samples D b ; The last three categories are used as incremental training samples D i The incremental learning stage belongs to small sample incremental learning, to be precise, 1-way-5-shot small sample learning. The training samples of this small sample learning are taken from the incremental training samples according to random indexes.

[0075] At the same time, the test set with a pitch angle of 15° in the MSTAR radar image data set is used as the validation set of this embodiment. Since the ability of the continuously evolving classifier to recognize SAR images is learned in stages, the validation set used in each stage is also different. The validation set used in the base class training stage is V b , the validation set used in the kth incremental learning stage is

[0076] In this embodiment, the deep residual network includes 4 residual blocks. The specific structure of the deep residual network based on the prototype idea is: the first convolutional layer → the first ReLU activation layer → the first residual block → the second residual block → the third residual block → the fourth residual block → the adaptive average pooling layer → the fully connected layer → the graph attention network → the cosine distance classifier.

[0077] Among them, the specific structure of each residual block is: second convolution layer → first batch normalization layer → second ReLU activation layer → third convolution layer → second batch normalization layer → third ReLU activation layer. In this embodiment, it is necessary to determine whether the number of input channels of the residual block is equal to the number of output channels. If they are equal, the input of the residual block and the input are passed through the second convolution layer → first batch normalization layer → second ReLU activation layer → third convolution layer → output of the second batch normalization layer, and the result of the addition is output after passing through the third ReLU activation layer. If they are not equal, the input of the residual block is passed through an additional convolution layer and an additional batch normalization layer (channel number conversion unit) and then passed through the second convolution layer → first batch normalization layer → second ReLU activation layer → third convolution layer → output of the second batch normalization layer, and the result of the addition is output through the third ReLU activation layer.

[0078] This is because the residual block is designed to require that its input and output are additive, so when the number of channels is different, the input will pass through a convolutional layer and a batch normalization layer to make its number of channels equal to the number of channels of the output at this time.

[0079] The training samples of the input deep residual network are SAR images with 3 channels and a size of 128×128. Each residual block contains two convolutional layers, and the number of output channels of the residual block is determined by the number of output channels of the second convolutional layer, that is, by the number of convolution kernels of the second convolutional layer; the number of convolution kernels of the second convolutional layer in the first residual block, the second residual block, the third residual block, and the fourth residual block are 32, 64, 128, and 256 respectively; the size of the convolution kernel inside the residual block is 3×3, and the size of the convolution kernel of the additional convolutional layer is 1×1; the number of convolution kernels of the first convolutional layer is 16, and the size is 3×3; the SAR image will become 2×2 feature information after passing through all residual blocks, and then pass through an adaptive average pooling layer with a set output of 1×1; the fully connected layer contains 256 nodes, which are linearly connected to 10 nodes.

[0080] In this embodiment, first, all base class training samples x are b and its corresponding classification label y b Send it into the deep residual network based on the prototype idea to get the class prototype prto of the base class b, and take another test sample as the validation set to prevent overfitting, and obtain a classifier that can recognize the base class (that is, the network model in the base class training phase). Secondly, according to the new class index, a small amount of incremental training samples x i and its corresponding classification label y i The network model is sent to the base class training phase, and the network model continues to learn the new class features and obtain the new class prototype prto i ,In this way, the continuously evolving classifier learns the ability to recognize new classes in the class incremental learning phase.

[0081] Specifically, the specific steps of base class training are:

[0082] 1a) Initialize the network parameters of the deep residual network based on the prototype idea and set the number of iterations in each stage T ≥ 100.

[0083] 1b) Extract the base class training sample x according to the pre-obtained index b and its corresponding classification label y b , the base class training sample x b The sample is sent to the deep residual network based on the prototype idea, and after passing through the convolution module and the deep residual network, the mapping f(x b ), where f(x b ) is a mapping expression with dimension N=256;

[0084] Among them, f() is the mapping function of the network, and the expression of the sample after mapping is:

[0085] encoder=f(x b ):x b ∈R 128×128 →f(x b )∈R 256 (1).

[0086] 1c) The fully connected layer is In the base class training phase, the 7 initialized class prototypes are stored in the weights of the fully connected layer as the weights of the fully connected layer. The class prototype calculation formula is as follows:

[0087]

[0088] in, is the kth prototype of the base class, is the k-th base class training set, for The total number of samples.

[0089] 1d) Multiplying the 256-dimensional mapping expression of the training sample by 7 weights can obtain the cosine similarity d(f(x b),prto b ), the calculation formula of cosine similarity is as follows:

[0090]

[0091] The larger the cosine similarity, the smaller the cosine distance, and the more relevant the two are. So according to the cosine similarity d(f(x b ),prto b ) can get the predicted classification result of the sample, and then combine it with the classification label y of the sample b , we can calculate the cross entropy loss function loss and accuracy acc;

[0092] Among them, the predicted classification results of the samples are:

[0093]

[0094] The training error calculation formula is as follows:

[0095]

[0096] 1e) According to the cross entropy loss function loss, back propagation updates the network parameters of the deep residual network based on the prototype idea, including the class prototype prto stored in the fully connected layer b In this way, the feature map of the base class training sample will be closer to its corresponding class prototype, thereby obtaining a clearer decision boundary between the base class training samples, completing the base class training stage, and obtaining the network model of the base class training stage.

[0097] Specifically, the specific steps of incremental learning are:

[0098] 2a) According to the new class index, a small number of incremental training samples x i and its corresponding classification label y i It is sent to the network model in the base class training phase and mapped into the feature space to obtain its mapping expression as f(x i ).

[0099] 2b) Take the average of the mapping expressions of samples of the same class and obtain the new class prototype prto in the feature space i , and pass the new class prototype to i As the 8th weight of the fully connected layer, the first 7 weights are the class prototypes in the base class training phase.

[0100] 2c) The network introduces a new parameter. The first 8 categories of samples in the training set (including 7 categories of training samples in the base class training stage and 1 category of incremental training samples in the incremental learning stage) need to be input into the network after the new parameters are introduced. The loss function is calculated based on the classification labels corresponding to the training samples, and then the 8 weights of the fully connected layer are updated by backpropagation.

[0101] 2d) When entering the second incremental learning stage, repeat steps 2a)-2c) to average the mapping of the new class samples to obtain the new class prototype as the 9th weight of the fully connected layer, and then use the first 9 classes of training set samples to update these 9 weights.

[0102] 2e) The same is true for the third incremental learning stage. Steps 2a)-2c) are repeated to average the mapping of the new class samples to obtain the new class prototype, which is used as the 10th weight of the fully connected layer. The first 10 classes of training set samples are then used to update these 10 weights to obtain a deep residual network based on the prototype idea that has been trained and learned.

[0103] In this embodiment, a deep residual network based on the prototype idea is used to train with MSTAR image samples in stages, and verified with a test set. In the base class training stage, a classifier with relatively good recognition performance can be obtained, but as small sample new class data is continuously fed into the network, that is, in the incremental learning stage, the performance of the classifier will decline, and this knowledge forgetting is bound to exist. Therefore, in this embodiment, the prototypes of the first 7 classes will also be used as the weights of the network's fully connected layer during base class training; in each class incremental learning stage, a new class prototype will be obtained, and the obtained class prototype will be fine-tuned. This process is equivalent to using training samples to calculate the forward propagation error in a graph attention network, and then back-propagating the error to update the weights of the network's fully connected layer, that is, fine-tuning the relative position of the class prototype, thereby obtaining a better classification boundary, allowing the continuously evolving classifier to obtain the ability to accurately identify new class samples and improve network recognition performance.

[0104] In the SAR image target recognition method based on small sample incremental learning of this embodiment, a deep residual network based on the prototype idea is designed. Even when the image samples are small, the network can automatically extract the features of the SAR image. Compared with the existing technology, the method of mapping the SAR image into the feature space can reduce the overfitting phenomenon caused by the lack of sample quantity to a certain extent, and improve the overall network's recognition ability for small sample images.

[0105] In the SAR image target recognition method based on small sample incremental learning of this embodiment, a continuously evolving classifier is designed to continue learning new classes in an already trained network. This setting is more in line with the need to continue learning small sample new class SAR image information in reality. In addition, a graph attention network is used to store various prototypes, and the distance between these prototypes is fine-tuned in the graph attention network to obtain a better classification boundary. This storage method is directly integrated in the fully connected layer of the network and is trained together with the network according to the training samples. It will have a better decision boundary than the network with fixed class prototypes, and improve the recognition accuracy.

[0106] Embodiment 2

[0107] This embodiment illustrates the effect of the SAR image target recognition method based on small sample incremental learning of the first embodiment through simulation experiments.

[0108] (1) Experimental conditions

[0109] The hardware platform of the simulation experiment of this embodiment is as follows: the GPU is NVIDIA GeForce RTX 2080Ti, 20 cores, main frequency is 2.2GHz, memory size is 128GB; the video memory size is 11GB.

[0110] The software platform of the simulation experiment in this embodiment is: the operating system is Windows 10.

[0111] The base class training samples of the simulation experiment of this embodiment are selected from 7 types of SAR images observed by the radar at a pitch angle of 17° in the MSTAR data set, namely BMP-2 (TG1), BTR-70 (TG2), T-72 (TG3), 2S1 (TG4), BRDM-2 (TG5), ZSU-234 (TG6) and ZIL-131 (TG10); the incremental training samples used in the incremental learning stage are the remaining three categories, namely BTR-60 (TG7), D-7 (TG8) and T-62 (TG9).

[0112] The test set samples are the test set in the MSTAR dataset with a pitch angle of 15°.

[0113] (2) Simulation content

[0114] In order to verify the small sample recognition capability of this embodiment, the simulation experiment sets two small sample incremental learning modes, 1-way-1-shot and 1-way-5-shot. The simulation results are shown in Table 1:

[0115] Table 1. Simulation results

[0116]

[0117]

[0118] It can be concluded from Table 1 that when the number of samples in the incremental learning stage is small, the continuously evolving classifier can also reach a higher level. From the average accuracy of the test, it can be seen that even if only one sample is used in each iteration, the average accuracy of recognition is only slightly lower than that of using five samples in each iteration. This shows that the deep residual network based on the prototype idea can overcome the overfitting phenomenon caused by small samples and can generate a better decision boundary for SAR image classification.

[0119] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of further restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements. Similar words such as "connect" or "connected" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0120] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A SAR image target recognition method based on small sample incremental learning, characterized in that: include: Step 1: Construct a deep residual network based on the prototype idea; Step 2: Using the base class data set and the incremental data set, the prototype-based deep residual network is respectively subjected to base class training and incremental learning; Step 3: Input the SAR image to be tested into the deep residual network based on the prototype idea that has been trained and learned to obtain the prediction classification result of the SAR image to be tested; The deep residual network based on the prototype idea includes a convolutional module, a deep residual network, an adaptive average pooling layer, a fully connected layer and a classifier connected in sequence, wherein: The convolution module includes a first convolution layer and a first ReLU activation layer connected in sequence; The deep residual network includes a plurality of residual blocks connected in sequence, and is used to extract image features of an input image; The fully connected layer is used to store class prototypes, and a graph attention network is integrated in the fully connected layer, and the graph attention network is used to adjust the position of the class prototype stored in the fully connected layer in the incremental learning stage; The classifier is a cosine classifier, which is used to output a predicted classification result; The residual block includes a second convolutional layer, a first batch of normalization layers, a second ReLU activation layer, a third convolutional layer, a second batch of normalization layers, and a third ReLU activation layer connected in sequence; The input of the residual block and the input of the residual block are sequentially passed through the second convolutional layer, the first batch of normalization layers, the second ReLU activation layer, the third convolutional layer, and the output of the second batch of normalization layers, and the result is output after passing through the third ReLU activation layer; The residual block further includes a channel number conversion unit, the channel number conversion unit is connected between an input end of the second convolutional layer and an output end of the second batch normalization layer, and the channel number conversion unit includes a convolutional layer and a batch normalization layer connected in sequence; When the number of input channels of the residual block is not equal to the number of channels of its output, the input of the residual block is converted by the channel number conversion unit to have the same number of channels as its output, and the result of adding the input of the residual block to the output of the second batch normalization layer is output after passing through the third ReLU activation layer.

2. The SAR image target recognition method based on small sample incremental learning according to claim 1 is characterized in that: The base dataset and the incremental dataset both include a plurality of SAR images of multiple categories with classification labels, and the incremental dataset is a 1-way-5-shot or 1-way-1-shot small sample dataset.

3. The SAR image target recognition method based on small sample incremental learning according to claim 1 is characterized in that: The step 2 comprises: Step 2.1: In the base class training stage, the base class data set is used to train and learn the deep residual network based on the prototype idea, and the network weights of the deep residual network based on the prototype idea are updated through back propagation to obtain the network model of the base class training stage; Step 2.2: In the incremental learning stage, the network model of the base class training stage is trained and learned using the incremental data set, and the weights of the fully connected layer are updated through back propagation to obtain a deep residual network based on the prototype idea that has been trained and learned.

4. The SAR image target recognition method based on small sample incremental learning according to claim 3 is characterized in that: The step 2.1 comprises: Step 1a: Initialize the network parameters of the deep residual network based on the prototype idea; Step 1b: obtaining base class training samples and their corresponding classification labels in the base class data set according to the index, inputting the base class training samples into the deep residual network based on the prototype idea, and obtaining the mapping of the base class training samples in the feature space after passing through the convolution module and the deep residual network; Step 1c: the mapping is input into the classifier after passing through the adaptive average pooling layer and the fully connected layer, and the classifier outputs the cosine similarity between the base class training sample and the class prototype; Step 1d: taking the class corresponding to the index with the largest cosine similarity as the predicted classification result of the base class training sample, and calculating the cross entropy loss function and accuracy according to the classification label corresponding to the base class training sample; Step 1e: According to the cross entropy loss function, the network weights of the deep residual network based on the prototype idea are updated through back propagation, and the mapping expression of the base class training samples in the feature space is taken as the class prototype, and stored as the weight of the fully connected layer to complete the base class training stage of the deep residual network based on the prototype idea and obtain the network model of the base class training stage.

5. The SAR image target recognition method based on small sample incremental learning according to claim 4 is characterized in that: The step 2.2 comprises: Step 2a: Obtain incremental training samples and their corresponding classification labels in the incremental data set, input the incremental training samples into the network model of the base class training stage, and obtain their mapping in the feature space; Step 2b: taking the average of the mapping expressions of samples of the same class, obtaining a new class prototype of the new class in the feature space, and using the new class prototype as the newly added weight of the fully connected layer; Step 2c: input the base class training samples and the incremental training samples into the network model obtained in step 2b, and calculate the cross entropy loss function according to the base class training samples and the incremental training samples and their corresponding classification labels; Step 2d: According to the cross entropy loss function, the weights of the fully connected layer are updated by back propagation to obtain a deep residual network based on the prototype idea that has been trained and learned.

Images