Shared Representation of SAR Target and Shadow Based on Multilayer Auto-encoder

Abstract

Automatic Target Recognition (ATR) of Synthetic Aperture Radar (SAR) images is investigated. A SAR feature extraction algorithm based on a multilayer auto-encoder is proposed. The method makes use of a probabilistic neural network and Restricted Boltzmann Machine (RBM) modeling probability distribution of the environment. Through the formation of a more expressive multilayer neural network, the deep learning model learns the shared representation of the target and its shadow outline reflecting the target shape characteristics. Targets are classified automatically through two recognition models. The experiment results based on the MSTAR verify the effectiveness of the proposed algorithm.