Multi-Scale Geometric Feature Optimization for Object Detection in SAR Images

Object detection in Synthetic Aperture Radar (SAR) images is a de facto need in military and civilian fields. The detector frequently encounter difficulties in capturing object geometric information due to multi-scale objects and unobvious edge features. To tackle this issue, we focus on cross-scale geometric information perception ability of SAR image object detectors. First, we construct a YOLOv8-n network based SAR object detection framework. Secondly, we propose a cross-space learning method to tackle short-range and long-range dependence, which could alleviate the small object false alarm problem. We propose a novel box regression loss that utilizes the minimum point distance, integrating the center point offset and width-height variation of horizontal bounding boxes to facilitate accelerated convergence. Comprehensive experiments conducted on the SSDD, MSTAR, and SAR-AIRcraft-1.0 datasets highlight the effectiveness and advantages of our approach. The mean average precision (mAP) on these datasets reaches 97.6%, 99.22%, and 81.23%, respectively, surpassing existing detection methods.