MsF-AT: A Study on Ship SAR Image Classification Based on Multi-Scale Feature and Attention Mechanism

Ship Synthetic Aperture Radar (SAR) images are a critical component of remote sensing imaging. Leveraging the all-weather and round-the-clock capabilities of radar detection, ship SAR has found extensive application in marine remote sensing monitoring and management. Despite achieving relatively high resolution at present, the Automatic Target Recognition (ATR) accuracy of ship SAR remains suboptimal due to limited image features and inherent SAR imaging characteristics, which affect its performance and broader adoption. In recent years, the rapid development of deep learning methods has significantly enhanced classification performance through innovative classifier design strategies and automatic feature extraction mechanisms. This paper proposes a novel deep neural network for classifying ship SAR images by integrating attention mechanisms and multi-scale feature fusion techniques within deep convolutional networks. The aim is to enhance feature channels and effectively leverage low-level features. Experimental results demonstrate that this model outperforms traditional methods, achieving 1.64% higher accuracy on three-category classification and 3.97% higher accuracy on six-category classification compared to the second-best method on the OpenSARShip dataset.