Abstract: Synthetic Aperture Radar (SAR) plays a vital role in ship detection due to its capability to capture high-resolution images in complex environments. However, existing detection models still suffer from false alarms and missed detections under such conditions. To address these challenges, this paper proposes a novel detection method named AABW-YOLO, based on an enhanced YOLOv5 framework. First, to improve multi-scale target detection, an Adaptive Kernel Convolution (AKConv) module is introduced for dynamic convolution kernel adjustment, and the original model’s “neck” is replaced with an Asymptotic Feature Pyramid Network (AFPN); second, the BiFormer attention mechanism is integrated into YOLOv5’s backbone to enhance the detection capability for small vessels in cluttered environments; third, a WIoU loss function is adopted to accelerate convergence and improve generalization. Evaluations on the SSDD and HRSID datasets demonstrate that the enhanced network achieves significant performance improvements, with AABW-YOLO attaining AP values of 97% and 83.7% on the two datasets, respectively, outperforming the baseline YOLOv5 and other state-of-the-art methods.