Abstract:

To address the issues of high model redundancy, computational complexity, and poor real-time performance in current sythetic aperture radar （SAR） image ship detection algorithms, this paper proposes a detection method of lightweight feature optimizetion based on EfficientNetv2 YOLO （you only look once） aimed at improving detection accuracy while reducing model complexity. It replaces YOLOv11 original backbone network with EfficientNetv2, effectively compressing parameters while preserving feature extraction capability. In the neck network, a dilated residual module and a spatial-channel reconstruction convolution module are introduced to enhance contextual awareness and fine-grained feature representation. A PIoUv2 localization loss function based on anchor box quality is designed to improve target regression accuracy. An appropriate pruning strategy is employed to streamline the model structure and enhance inference efficiency. Experimental results on two benchmark SAR ship detection datasets, HRSID（SAR Ship Detection Dataset, SSDD） and SSDD（High Resdution SAR Image Dataset, HRSID）, demonstrate that the proposed method significantly reduces computational overhead and inference time while maintaining high detection accuracy. The method exhibits strong adaptability across diverse scenarios and holds substantial potential for practical applications, providing a viable solution for lightweight platforms in radar image detection.

Key words: synthetic aperture radar （SAR） image, lightweight, feature extraction, ship detection, target detection