Hybrid Conv-Attention Networks for Synthetic Aperture Radar Imagery-Based Target Recognition

Abstract

In this study, we propose hybrid conv-attention networks that combine convolutional neural networks (CNNs) and transformers to recognize targets from synthetic aperture radar (SAR) images automatically. The proposed model is designed to obtain robust features from global and local patterns in the SAR image, utilizing the weights of a pre-trained backbone model with self-attention structures. Furthermore, we adopted pre-processing and training methods optimized for transfer learning to enhance performance. By comparing and analyzing the performance between the proposed model and conventional models using the OpenSARShip and MSTAR dataset, we found that our system significantly outperforms conventional approaches, with a performance improvement of 24.06%. This considerable enhancement is attributed to the ability of the model to leverage the 2D kernel-based approach of CNNs and the sequence vector-based approach of transformers, offering a comprehensive method for SAR image target recognition.

Keywords

• Synthetic aperture radar (SAR)
• target recognition
• deep learning (DL)
• transfer learning
• convolutional neural networks (CNNs)
• transformers