CN112446357A - SAR automatic target recognition method based on capsule network - Google Patents

Abstract

The invention discloses an SAR automatic target recognition method based on a capsule network, and belongs to the field of radar target recognition. The method comprises the following main processes: firstly, cutting an original SAR image, then simply convolving the cut image, then extracting multi-scale features by using convolution kernels with different expansion rates, then enhancing important features by using a self-adaptive feature refinement module, fusing the enhanced multi-scale features by a pixel-by-pixel fusion strategy, then inputting the features into a network layer based on a capsule unit for more abstract feature learning and reserving the spatial relationship among the features, and finally inputting the features output by an encoder network into a decoder network consisting of four transposed convolution layers for SAR target reconstruction to improve the learning capability of the encoder; and the SAR target identification result is output at the last layer of the encoder network. Compared with the existing deep convolutional neural network algorithm, the method has higher precision.

Description

SAR automatic target recognition method based on capsule network

Technical Field

The invention is applied to the field of Synthetic Aperture Radar (SAR) automatic target identification, and particularly relates to an SAR automatic target identification method based on deep learning.

Background

Synthetic aperture radars have been widely used in the fields of geological exploration, environmental monitoring, military target detection and identification, etc. due to their all-weather, all-time, high-resolution, etc. characteristics. Automatic Target Recognition (ATR) is one of the important applications for SAR image interpretation.

In recent years, with the continuous development of deep learning technology, in the field of automatic target recognition of SAR, various ATR algorithms based on deep learning models have been proposed and achieve better recognition performance than the conventional method under some Standard conditions (SOC). However, under Extended Operating Conditions (EOCs), such as in a noisy environment, various complex noises can seriously affect the image feature extraction; in most real SAR scenes, it is difficult to collect a large number of training samples, and under the condition that the training samples are insufficient, a classification model is easy to be over-fitted; partial target occlusion and camouflage is common in combat scenarios, and it is challenging to extract robust discriminating features from partial target occlusions. In view of the above problems, the present invention provides a method for identifying an SAR target based on a convolutional capsule network. In particular, the method comprises two sub-networks: an encoder network and a decoder network. The encoder network may extract robust features from the SAR image. The decoder network may encourage the encoder network to learn the authentication features.

Disclosure of Invention

Aiming at the problems in SAR ATR, the invention provides a method based on a convolution capsule network to realize high-precision SAR automatic target recognition.

The technical scheme adopted by the invention for solving the problems is as follows: a deep classification model is designed, which is composed of an encoder network and a decoder network. Specifically, in consideration of the fact that the depth model can learn local features such as textures and shapes in a lower-layer network, in order to relieve the influence of noise on SAR target identification, the method extracts multi-scale features through convolution of a plurality of holes in the lower-layer network of an encoder; considering that the multi-scale features possibly contain redundant information, embedding a feature self-adaptive refining module on a multi-scale feature channel to self-adaptively enhance useful features and inhibit useless features; a top-level network considering depth models can learn the structural characteristics of image features, and therefore, two capsule-unit-based feature structure-preserving layers are deployed at the top level of the encoder network to learn more abstract features and preserve spatial relationships between different features. The decoder network composed of multiple layers of transposed convolutional layers can realize image reconstruction to further improve the learning capability of the encoder network.

The technical scheme of the invention is a capsule network-based SAR automatic target recognition method, which comprises the following steps:

step 1: cutting the acquired SAR image into 64 × 64 slices to reduce the influence of redundant background on feature extraction;

step 2: carrying out preliminary feature learning through a convolution layer with the kernel size of 3 x 3 to obtain 32 feature maps of 60 x 60;

and step 3: extracting features of different scales by using 3 convolution kernels with different expansion rates for each feature map, wherein the size of the convolution kernels is 5 multiplied by 5, and the expansion rates are 2, 3 and 5 respectively;

and 4, step 4: respectively passing the features of each scale through an adaptive feature refinement module, wherein the processing method of the adaptive feature refinement module comprises the following steps:

Q＝σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

S＝σ(MLPConv_7×7([AvgPool(Fc)；MaxPool(Fc)]))

wherein F represents the input of the adaptive feature refinement module which is one of the scale features obtained in the step 3, AvgPool and MaxPool respectively represent average pooling and maximum pooling, MLP represents a perceptron including an implied layer, sigma is a sigmoid function, Q represents the weight of a corresponding channel,

representing a pixel-by-pixel fusion operation, Conv_7×7Representing a 7 × 7 convolutional layer, S represents a spatial weight, and Fout is a final output through an adaptive feature refinement module;

and 5: then, the identification features obtained in the step 4 are sampled to the same size, and the sampled features are fused to fuse a feature F_MeltComprises the following steps:

F_melt＝T(Fout₁)+T(Fout₂)+T(Fout₃)

Where T represents the upsampling operation, Fout₁，Fout₂，Fout₃Respectively representing the 3 scales of the identification features obtained by the step 4;

step 6: inputting the fused multi-scale features into a capsule unit-based network layer;

step 6.1: fusing the characteristics F obtained in the step 5_MeltPerforming convolution operation and converting the operation into vector characteristics Cap;

step 6.2: performing nonlinear conversion on the characteristics of the Cap by adopting the following formula;

wherein, Cap_iRepresenting the ith vector feature, u_iIs the result of vector feature non-linearisation;

step 6.3: learning a characteristic space relation by adopting the following formula;

u_j|i＝W_ij·u_i

wherein, W_ijFor feature mapping matrices learned by back propagation algorithms, c_ijIs the capsule coefficient obtained by the dynamic routing algorithm;

step 6.4: the characteristics s obtained in step 6.3_jOutputting through an SAR capsule layer;

and 7: inputting the features output in step 6 into a decoder network consisting of four transposed convolutional layers, wherein the specific parameters of the transposed convolutional layers are as follows: a first layer: the convolution kernel size is 7 × 7, the convolution step is 1, and the padding is 0; a second layer: convolution kernel size 5 × 5, convolution step size 2, padding 1; and a third layer: the convolution kernel size is 5 × 5, the convolution step size is 2, and the padding is 1; a fourth layer: the convolution kernel size is 4 × 4, the convolution step size is 2, and the padding is 0.

The invention extracts the multi-scale features by introducing the hole convolution, and can relieve the influence of noise on target identification. In addition, considering that some features with small information amount or useless identification are probably contained in the multi-scale features, the target identification performance can be further improved by embedding an adaptive feature refining module into a multi-scale channel for adaptive feature weighting. Considering that training a classification model based on a convolutional neural network requires a large number of training samples, however, capturing a large number of SAR images is difficult in most cases. In order to solve the problem, the excessive dependence of the model on the training sample size is relieved by designing a characteristic space relation retaining layer by utilizing capsule units.

Drawings

FIG. 1 is a network architecture of a convolutional capsule network;

FIG. 2 architecture of a feature refinement module;

FIG. 3 shows the recognition result of the method proposed in the present invention under the condition of small sample training number;

FIG. 4 shows the recognition results of the method proposed in the present invention under the condition of partial target occlusion of different degrees.

Detailed Description

Hereinafter, a detailed description will be given of an embodiment of the present disclosure in order to better embody the technical points of the present disclosure. The invention relates to an SAR target recognition method based on a convolution capsule network, and each step is implemented in the following mode.

Step 1: cutting the acquired SAR image into 64 × 64 slices to reduce the influence of redundant background on feature extraction;

step 2: carrying out preliminary feature learning through a convolution layer with the kernel size of 3 x 3 to obtain 32 feature maps of 60 x 60;

and step 3: features of different scales are extracted using a plurality of convolution kernels of different expansion ratios. The expansion convolution used in the invention is used for extracting features of different scales, specifically, the convolution kernel size is 5 multiplied by 5, and the expansion rates are 2, 3 and 5 respectively;

and 4, step 4: the multi-scale features are input to an adaptive feature refinement module that adaptively enhances useful features and suppresses useless information. Order to

Is a multi-scale feature, the channel weight of the adaptive feature refinement module

And spatial weight

The calculation process is as follows:

Q＝σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

S＝σ(MLPConv_7×7([AvgPool(Fc)；MaxPool(Fc)]))

wherein Fout is the discriminative features learned by the adaptive feature refinement module, σ is the sigmoid function, MLP is derived from a perceptron comprising a hidden layer,

representing a pixel-by-pixel fusion operation, AvgPool and MaxPool representing average pooling and maximum pooling, Conv, respectively_7×7Representing 7 × 7 convolutional layers.

And 5: fusing the refined multi-scale features through a pixel-by-pixel fusion strategy, wherein the fusion features are as follows:

F_melt＝T(Fout₁)+T(Fout₂)+T(Fout₃)

Where T represents the upsampling operation, Fout₁，Fout₂，Fout₃Respectively representing the 3 multi-scale identification characteristics obtained in the step 4;

step 6: inputting the fused multi-scale features into a capsule unit-based network layer;

step 6.1: fusing the characteristics F obtained in the step 5_MeltPerforming convolution operation and converting the operation into vector characteristics Cap;

step 6.2: performing nonlinear conversion on the characteristics of the Cap by adopting the following formula;

wherein, Cap_iRepresenting the ith vector feature, u_iOutputting after the vector characteristic is nonlinear;

step 6.3: learning a characteristic space relation by adopting the following formula;

u_j|i＝W_ij·u_i

wherein, W_ijFor feature mapping matrices learned by back propagation algorithms, c_ijThe capsule coefficient is obtained through a dynamic routing algorithm;

step 6.4: step 6.3 features s_jOutputting through an SAR capsule layer;

and 7: features output by the encoder network are input to a decoder network consisting of four transposed convolutional layers in order to facilitate the encoder network to learn more discriminating features from the SAR image.

In the embodiment, the method adopts the MSTAR data set of the public standard for verification, and can complete high-precision SAR target identification compared with other deep networks. Four different verification approaches were devised:

(1) verifying the identification capability of the convolution capsule network under standard operating conditions, and adopting ten types of foundation military targets: BMP2, BRDM _2, BTR70, BTR60, T72, 2S1, D7, T62, ZIL131 and ZSU23_ 4. The SAR image collected under the pitch angle of 17 degrees is used as a training set, the data collected under the pitch angle of 15 degrees is used as a test set, and the experimental data are shown in table 1. In order to show that the method provided by the invention has high discrimination capability, three ATR algorithms based on deep learning are compared in experiments, and the algorithms are DCNN, A-ConvNets and MFCNNs respectively. The results of the experiment are shown in table 2. As can be seen from table 2, the model proposed by the present invention can obtain the optimal recognition performance compared to the comparative method.

(2) And verifying the identification capability of the convolution capsule network provided by the invention under the noise interference condition. Training data as in experiment (1), we performed noise contamination to different degrees on the test images using a simulation method. Specifically, the original element values of different ratios in the test image are replaced with random numbers between (0, 1). In this experiment, the noise pollution ratio was changed from 1% to 15%, and the experimental results are shown in table 3. From experimental results, along with the improvement of the noise pollution level, the method provided by the invention has higher identification precision than that of a comparison method. Experimental results show that the method has strong noise robustness.

(3) And verifying the identification capability of the method provided by the invention under the condition of insufficient training samples. Specifically, a small number of samples are randomly drawn from the original training set to simulate a limited training sample scenario, and then the network is trained using these limited samples, and the experimental result is shown in fig. 3. From the experimental results, the method provided by the invention is still superior to the comparison method under the condition of insufficient training samples. Particularly, when the training sample only has 10% of the original training data, the method provided by the invention can still achieve the recognition accuracy of more than 80%, and the recognition accuracy of other methods is lower than 80%. The experimental result shows that the method provided by the invention still has advantages over the existing methods in the scene of limited training samples.

(4) The identification capability of the method provided by the invention is verified under the condition that part of the target is shielded. In this experiment, we used a random erasure method to simulate part of the target occlusion data. In the experiment, the results of the recognition capability of different networks under different levels of shielding of the target are shown in fig. 4. From experimental results, the recognition ability of the proposed method of the present invention is always superior to the comparative method at different occlusion levels. In a real military target battle scene, target shielding and disguising are very common, and the experimental result shows that the research content of the invention has potential application value.

Table 1 description of the data set used in the experiment

Table 2 overall recognition accuracy for four network architectures under standard operating conditions

Network name	DCNN	MFCNNs	A-ConvNets	Algorithm of the invention
					Percent identification (%)	92.30	95.52	95.27	99.18

Table 3 overall recognition accuracy of four network architectures under noisy interference conditions

Claims (1)

1. A capsule network-based SAR automatic target recognition method comprises the following steps:

step 1: cutting the acquired SAR image into 64 × 64 slices to reduce the influence of redundant background on feature extraction;

step 2: carrying out preliminary feature learning through a convolution layer with the kernel size of 3 x 3 to obtain 32 feature maps of 60 x 60;

and step 3: extracting features of different scales by using 3 convolution kernels with different expansion rates for each feature map, wherein the size of the convolution kernels is 5 multiplied by 5, and the expansion rates are 2, 3 and 5 respectively;

and 4, step 4: respectively passing the features of each scale through an adaptive feature refinement module, wherein the processing method of the adaptive feature refinement module comprises the following steps:

Q＝σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

S＝σ(MLPConv_7×7([AvgPool(Fc)；MaxPool(Fc)]))

wherein F represents the input of the adaptive feature refinement module which is one of the scale features obtained in the step 3, AvgPool and MaxPool respectively represent average pooling and maximum pooling, MLP represents a perceptron including an implied layer, sigma is a sigmoid function, Q represents the weight of a corresponding channel,

representing a pixel-by-pixel fusion operation, Conv_7×7Representing a 7 × 7 convolutional layer, S represents a spatial weight, and Fout is a final output through an adaptive feature refinement module;

and 5: then, the identification features obtained in the step 4 are sampled to the same size, and the sampled features are fused to fuse a feature F_MeltComprises the following steps:

F_melt＝T(Fout₁)+T(Fout₂)+T(Fout₃)

Where T represents the upsampling operation, Fout₁，Fout₂，Fout₃Respectively representing the 3 scales of the identification features obtained by the step 4;

step 6: inputting the fused multi-scale features into a capsule unit-based network layer;

step 6.1: fusing the characteristics F obtained in the step 5_MeltPerforming convolution operation and converting the operation into vector characteristics Cap;

step 6.2: performing nonlinear conversion on the characteristics of the Cap by adopting the following formula;

wherein, Cap_iRepresenting the ith vector feature, u_iIs the result of vector feature non-linearisation;

step 6.3: learning a characteristic space relation by adopting the following formula;

u_j|i＝W_ij·u_i

wherein, W_ijFor feature mapping matrices learned by back propagation algorithms, c_ijIs the capsule coefficient obtained by the dynamic routing algorithm;

step 6.4: the characteristics s obtained in step 6.3_jOutputting through an SAR capsule layer;

and 7: inputting the features output in step 6 into a decoder network consisting of four transposed convolutional layers, wherein the specific parameters of the transposed convolutional layers are as follows: a first layer: the convolution kernel size is 7 × 7, the convolution step is 1, and the padding is 0; a second layer: convolution kernel size 5 × 5, convolution step size 2, padding 1; and a third layer: the convolution kernel size is 5 × 5, the convolution step size is 2, and the padding is 1; a fourth layer: the convolution kernel size is 4 × 4, the convolution step size is 2, and the padding is 0.

Images