SAR ATR for Limited Training Data Using DS-AE Network

Abstract

Although automatic target recognition (ATR) with synthetic aperture radar (SAR) images has been one of the most important research topics, there is an inherent problem of performance degradation when the number of labeled SAR target images for training a classifier is limited. To address this problem, this article proposes a double squeeze-adaptive excitation (DS-AE) network where new channel attention modules are inserted into the convolutional neural network (CNN) with a modified ResNet18 architecture. Based on the squeeze-excitation (SE) network that employs a representative channel attention mechanism, the squeeze operation of the DS-AE network is carried out by additional fully connected layers to prevent drastic loss in the original channel information. Then, the subsequent excitation operation is performed by a new activation function, called the parametric sigmoid, to improve the adaptivity of selective emphasis of the useful channel information. Using the public SAR target dataset, the recognition rates from different network structures are compared by reducing the number of training images. The analysis results and performance comparison demonstrate that the DS-AE network showed much more improved SAR target recognition performances for small training datasets in relation to the CNN without channel attention modules and with the conventional SE channel attention modules.

Keywords: automatic target recognition (ATR); channel attention; convolutional neural network (CNN); deep learning; double-squeeze-adaptive-excitation network; limited labeled data; synthetic aperture radar (SAR).

Figure 1

Structure of the ResNet18 modified for SAR-ATR application. The above ResNet18 has 18 layers with learnable parameters and 8 skip connections, identical to the original ResNet18. The main modifications are: the input size, the receptive field size of the first convolutional layer, the number of nodes of the last fully connected layer, and the adoption of the pre-activation structure.

Figure 2

Structure of the SE channel attention module. The input feature map is squeezed and recalibrated via the global average pooling and two fully connected layers. Each fully connected layer is followed by the activation functions, the ReLU and the sigmoid.

Figure 3

Structure of the channel attention module of the DS-AE network. It has the double squeeze structure implemented by four sequential fully connected layers where each one is activated by the ReLU except for the last one followed by the parametric sigmoid with more adaptivity compared to the original sigmoid.

Figure 4

Shape of the parametric sigmoid (a) for different values of a and fixed b as 1, and (b) for different values of b and fixed a as 1. Note that both a and b are adaptively determined during the network learning process. When both a and b are 1, the parametric sigmoid is degenerated into the original sigmoid.

Figure 5

Optical and SAR images of 10 targets in the MSTAR dataset.

Figure 6

SAR target images and class activation maps of T72 (1st row) and BMP2 (2nd row).

Figure 7

16 strongest channel-wise activation maps of T72 SAR target image. (a) Activation maps from the basic CNN; (b) activation maps from the SE network; (c) activation maps from the DS-AE network.

Figure 8

16 strongest channel-wise activation maps of BMP2 SAR target image. (a) Activation maps from the basic CNN; (b) activation maps from the SE network; (c) activation maps from the DS-AE network.

Figure 9

Shapes of the original sigmoid and the parametric sigmoids for the test image of T72 input to the DS-AE network learned by 25% of MSTAR training images. The red and blue correspond to the parametric sigmoids of channels with strong and weak responses, respectively. (a) Original sigmoid; (b) parametric sigmoids of stage 1 in the basic CNN; (c) parametric sigmoids of stage 2 in the basic CNN; (d) parametric sigmoids of stage 3 in the basic CNN.

Figure 10

Shapes of the original sigmoid and the parametric sigmoids for the test image of BMP2 input to the DS-AE network learned by 25% of MSTAR training images. The red and blue correspond to the parametric sigmoids of channels with strong and weak responses, respectively. (a) Original sigmoid; (b) parametric sigmoids of stage 1 in the basic CNN; (c) parametric sigmoids of stage 2 in the basic CNN; (d) parametric sigmoids of stage 3 in the basic CNN.

Figure 11

Channel excitation vectors from the DS-AE network (blue line) and the SE network (red line) for stages 1, 2, and 3. (a) Channel excitation vectors at the output of stage 1 before excitation; (b) channel excitation vectors at the output of stage 1 after excitation; (c) channel excitation vectors at the output of stage 2 before excitation; (d) channel excitation vectors at the output of stage 2 after excitation; (e) channel excitation vectors at the output of stage 3 before excitation; (f) channel excitation vectors at the output of stage 3 after excitation.