SAR Reconnaissance: Unveiling Target Recognition Via Pretext and Downstream Methodologies

In this work, we propose a novel deep neural network for Automatic Target Recognition (ATR) in Synthetic Aperture Radar (SAR) images. Due to limited labeled data, the network architecture is designed with combination of pretext followed by downstream models. A pretext model is specifically designed for multi-scale feature integration and orientation estimation. Pretext and downstream models are trained independently and the knowledge is seamlessly transferred from pretext to downstream. The combination of multi-scale features and the orientation estimation in the pretext enhances the downstream model’s robustness by providing the fine details obtained at different scales of the image and supplying orientation information for precise target recognition and classification. The efficacy of the proposed model is verified by experimenting on the benchmark Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset achieving an accuracy of 99.77%. This contribution marks a substantial advancement in SAR technology, particularly in refining target recognition techniques.