BESW-YOLO: A Lightweight SAR Image Detection Model Based on YOLOv8n for Complex Scenarios

Abstract

Synthetic Aperture Radar (SAR) is a vital technology for ship detection due to its ability to capture high-resolution remote sensing images. However, traditional detection methods often suffer from false alarms and missed detections. Additionally, many current approaches prioritize detection accuracy while overlooking model size. To address these challenges, this paper proposes BESW-YOLO, a lightweight multi-scale ship detection model built upon the YOLOv8n architecture. Firstly, we introduce a novel lightweight feature pyramid network, Bidirectional and Multi-scale Attention Feature Pyramid Network (BiMAFPN), which effectively enhances the fusion of features across different scales. Secondly, Efficient Multi-Scale Convolution (EMSC) is introduced, which is combined with the C2f module in the YOLO model to form a new module, EMSC-C2f. This combination reduces the model parameters while improving feature extraction capabilities. Thirdly, to further optimize the model's multi-scale detection performance, we introduce a simple and efficient attention mechanism (SimAM), which enables adaptive weighting to emphasize target regions. Finally, the inner wise intersection over union loss function (Inner-WIoU) is introduced, which accelerates the model's convergence speed and enhances its generalization capability. The proposed BESW-YOLO model was evaluated using the SSDD, HRSID and SAR Ship dataset. Experimental results show that BESW-YOLO model achieves Average Precision (AP) values of 97.3%, 90.0% and 90.3% on the SSDD, HRSID and SAR Ship dataset, respectively, with only 1.7 M model parameters. It outperforms the baseline YOLOv8n in terms of both accuracy and model size. Compared to other mainstream models, BESW-YOLO delivers superior detection performance with significantly fewer parameters. These results confirm that BESW-YOLO is a lightweight and efficient detection model.

Topics & Concepts