Multi-scale Convolution and Feature-weighting Network for Radar Target Recognition

Target recognition is one of the most significant applications of synthetic aperture radar (SAR). However, satisfactory results are impractical to achieve by human effort alone due to the continual and rapid growth of the quantity of radar data. In view of the great success of convolutional neural networks (CNNs) in optical image classification tasks, in this paper, we apply a modified CNN to improve the classification accuracy. Instead of simply stacking several convolutional, pooling and activation layers to build a structure, a module that groups three different forms of convolution is designed to improve the feature extraction ability. Considering the complexity of SAR image composition, targets are not adequately described with single-scale feature maps. In addition, to utilize the correlation of features, an extra module is designed to measure the weights of the features and preprocess the input of the next stage. Experiments are performed on a moving and stationary target acquisition and recognition dataset. The proposed method achieves an average accuracy of 98% and a maximum accuracy of 99.67%, which demonstrates its efficiency compared with existing methods.