A concurrent and hierarchy target learning architecture for classification in SAR application

Abstract

This article discusses the issue of Automatic Target Recognition (ATR) on Synthetic Aperture Radar (SAR) images. Through learning the hierarchy of features automatically from a massive amount of training data, learning networks such as Convolutional Neural Networks (CNN) has recently achieved state-of-the-art results in many tasks. To extract better features about SAR targets, and to obtain better accuracies, a new framework is proposed: First, three CNN models based on different convolution and pooling kernel sizes are proposed. Second, they are applied simultaneously on the SAR images to generate image features via extracting CNN features from different layers in two scenarios. In the first scenario, the activation vectors obtained from fully connected layers are considered as the final image features; in the second scenario, dense features are extracted from the last convolutional layer and then encoded into global image features through one of the commonly used feature coding approaches, which is Fisher Vectors (FVs). Finally, different combination and fusion approaches between the two sets of experiments are considered to construct the final representation of the SAR images for final classification. Extensive experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset are conducted. Experimental results prove the capability of the proposed method, as compared to several state-of-the-art methods.