Enhanced YOLOv9-RDA for target detection in sonar images

Forward-looking sonar is critical for underwater target detection but is often hindered by image degradation and strict real-time processing demands. To address this, we propose an improved YOLOv9-RDA method for small target detection in complex environments. We introduce a lightweight RepViT architecture into the backbone, effectively reducing model depth and parameters while significantly improving inference speed without compromising accuracy. Furthermore, a DySAU module is designed to enhance background modeling and detail reconstruction, while the LSK attention mechanism is integrated to optimize the detection accuracy of small-scale targets. Experimental results show that YOLOv9-RDA achieves a 3.2% increase in mean Average Precision (mAP) and a 23% reduction in parameters, demonstrating excellent real-time performance. These improvements provide robust support for detection equipment on autonomous underwater vehicles (AUVs) and marine platforms.