Automatic SAR Target Recognition Based on Two-Dimensional Locality-Preserved Maximum Information Projection

Synthetic Aperture Radar (SAR) has been widely used in military and civil domains, while SAR Automatic Target Recognition (SAR ATR) poses a great challenge for researchers to meet real application demands. It is widely agreed that the lack of effective features is the bottleneck in advance of SAR ATR. In this paper, Locality-Preserved Maximum Information Projection (LPMIP) is used to extract features from the SAR images for ATR, where LPMIP is improved in three ways. 1) It is extended from one-dimensional version to two-dimensional one so that the structure information of SAR images can be better preserved. 2) The labels of samples are taken into account which makes the feature extraction method a supervised learning process. 3) To better balance the global and local structures, a two-stage framework is adopted, the presented 2D-LPMIP is executed following 2D-PCA to extract features of SAR images. ATR experiments on MSTAR databases achieved recognition accuracy of 97.44% with 698 training samples and 1662 testing samples. Experiment results testify the effectiveness of the proposed SAR image feature extraction method.