An automatic target recognition algorithm for SAR image based on improved convolution neural network

摘要

An Improved Convolution Neutral Network (CNN) model is proposed in view of the characteristics of Synthetic Aperture Radar (SAR) image. This model, denoted as Q-Net, consists of the convolution layer, sub-sampling layer, fully-connected layer and other basic structures. The error back-propagation algorithm is utilized to train the network parameters. Q-Net realizes automatic targets recognition in SAR images, and uses the Moving and Stationary Target Acquisition and Recognition (MSTAR) radar image database to train and test the model. In the case of merely cropping original images, Q-Net hits recognition accuracy of 97.58% in classifying three kinds of targets and 97.32% in ten kinds respectively. Experimental results show that compared with other methods, Q-Net not only achieves target recognition with higher accuracy, but also features less data and faster convergence, which can be a reference for the practical application.