SG-YOLO: Gain-Safe and Context-Aware Feature Enhancement for Ship Detection in SAR Images
Abstract
Small and weak ship signatures, strong background scattering, and localization sensitivity remain major challenges in synthetic aperture radar (SAR) ship detection. This letter proposes SG-YOLO, an SAR ship detector that combines adaptive frequency encoding, gain-safe feature enhancement, context-aware modulation, and localization-sensitive regression. The proposed adaptive cascaded wavelet feature extraction (ACWFE) block decomposes the low-frequency branch and combines learnable subband scaling and cross-level reconstruction before early spatial reduction. In the backbone, stabilized residual learning C3k2 with Star (SRLC3k2-Star) operation introduces multiplicative feature interaction through a zero-initialized bounded residual gain, enabling progressive enhancement of inherited representations. Large selective kernel attention-enhanced C3k2 (LSKAC3k2) further applies large-receptive-field spatial modulation at the fusion stages to improve target–background discrimination, while Inner-WIoU3 enhances regression performance for compact targets. Experiments on HRSID, SSDD, and LS-SSDD-v1.0 demonstrate absolute mAP ${}_{\mathrm {50:95}}$ improvements of 4.27%, 2.11%, and 4.45% over YOLO11s, together with recall gains of 4.37%, 1.67%, and 1.10%, respectively. Additional inshore-offshore and small-target evaluations further confirm the effectiveness of the SG-YOLO under challenging SAR scenes.