Abstract
Acoustic shadows, reverberation, and complex seabed conditions severely hinder shipwreck detection in side-scan sonar (SSS) imagery. We propose Deep Feature Selective Enhancement YOLO (DFSE-YOLO), a lightweight architecture designed to resolve the accuracy-efficiency bottleneck. Unlike traditional additive improvements that effectively increase computational burden, our approach integrates a threshold-based selective strategy that functions as a Channel-depth dependent selective strategy. This mechanism strategically allocates processing resources by enhancing semantic-rich deep features with adaptive residual connections and channel-aware depthwise convolutions, while allowing shallow features to bypass heavy computation. Evaluations on the AI4Shipwrecks dataset demonstrate that DFSE-YOLO achieves a mean Average Precision (mAP50) of 0.7551, surpassing YOLOv8 and YOLOv11 by 5.5% and 4.7%, respectively. With an inference speed of 40.1 FPS and a compact 5.25 MB footprint, the model is ideal for real-time deployment on autonomous underwater vehicles (AUVs). These results validate that incorporating acoustic-aware feature selection into lightweight architectures can effectively resolve the accuracy-efficiency bottleneck, offering a robust solution for real-world deployment on autonomous underwater vehicles. The code is publicly available at https://github.com/gmgslinyu/YOLOv11-improved4AI4ShipwrecksOD.