Target Recognition Method for High Resolution SAR Images Based on Improved Convolutional Neural Network

Deep Convolutional Neural Network (CNN) has obtained state-of-the-art accuracy in many image recognition tasks. It can learn hierarchical features from massive training data automatically. Since the number of SAR images is limited, using traditional CNN in SAR target recognition will yield severe overfitting. This paper proposes an improved CNN algorithm for high resolution SAR image target recognition. The CNN algorithm is trained by images with target rotation, target translation and random noise in target. With these training data, the system should be more robust and insensitive to these target transformations. During the training, a few strategies such as L2 regularization, batch normalization and dropout are investigated to restrain overfitting. Experimental results on Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset show that the proposed method could achieve high accuracy and be more robust.