SIAMESE NEURAL NETWORK FOR AUTOMATIC TARGET RECOGNITION USING SYNTHETIC APERTURE RADAR

Automatic Target Recognition (ATR) is an interest problem in various application fields (security, surveillance, automotive, environment, medicine, communications, remote sensing, ... ). Thus, SAR (Synthetic Aperture Radar) and ISAR (Inverse Synthetic Aperture Radar) radar images provide rich visual information about the observed radar target. From these radar images, several methods have been proposed to meet the expected requirements in several application domains, including target recognition, which is one of the main issues addressed in the present work. Traditional standard image classification techniques are not suitable for efficient classification of SAR images due to the limited data available in some classes (unbalanced data). To solve these problems, we introduce a deep learning model, the Siamese network with multi-class classification, built from a pre-trained model to improve the model performances on unbalanced classes. To evaluate the proposed method, the MSTAR dataset is used. The proposed method improves the recognition rate from 95,18% to 97.16%.