YOLO-RC: SAR Ship Detection Guided by Characteristics of Range-Compressed Domain
Xiangdong Tan, Xiangguang Leng, Ru Luo, Zhongzhen Sun, Kefeng Ji, Gangyao Kuang
Abstract
Conventional ship detection using synthetic aperture radar (SAR) is typically limited to fully-focused SAR images, limiting the development of real-time SAR ship detection. Ship detection in the SAR range-compressed domain holds significant real-time potential as it obviates the need for complete imaging and reduces data transmission. However, range-compressed data are solely compressed in range, resulting in a defocused representation in azimuth, which differs from SAR images. The previously proposed methods often fail to effectively incorporate the characteristics of range-compressed domain. In light of this circumstance, we propose an SAR ship detection network, YOLO-range compressed (YOLO-RC), which utilizes amplitude gradient and geometric scale characteristics in the range-compressed domain for improved performance. In YOLO-RC, amplitude gradient guided feature extraction module is specifically designed to leverage the different gradient variation trends of ship in both the range and azimuth dimensions. Moreover, we incorporate a large receptive field pyramid head, employing a pyramid-like structure to enhance receptive field and achieve more precise fitting of ship geometry for improved detection capability. Considering the scarcity of range-compressed ship samples, we conduct experiments using a publicly available self-built dataset. Experimental results on the dataset demonstrate that the proposed network achieves an F<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub>-score of 83.78% and an average precision of 84.09%, outperforming most existing SAR ship detection methods with better detection capability in SAR range-compressed domain.