Lightweight sonar target detection model for unmanned underwater vehicle via multi-scale feature fusion and acoustic perception

To address the urgent need for efficient and accurate perceptual capabilities in unmanned underwater vehicle (UUV)-based information countermeasure systems, this paper proposes a novel lightweight and edge-aware detection framework for forward-looking sonar images. Despite the advances in underwater target detection, existing models are often hindered by high computational costs, suboptimal boundary localization, and the lack of specialized sonar datasets. In this work, we propose Lightweight-YOLOv8, an object detection model tailored for sonar applications, and constructs a dedicated underwater sonar image (USI) dataset to support experimental validation. Specifically, we design a lightweight C2F module integrating inverted residual blocks to achieve a balance between detection accuracy and inference speed. To improve target representation, we propose a multi-scale and multi-receptive field feature fusion strategy. Furthermore, an edge-aware CIoU loss function is introduced by incorporating Canny edge detection to enhance boundary precision in sonar scenes. Experimental results on the USI dataset show that our model achieves a precision of 98.45%, surpassing YOLOv7 and YOLOv8 by 9.73% and 4.99%, respectively. Ablation studies confirm the contribution of each component. The proposed framework offers a promising solution for distributed, real-time acoustic perception in intelligent UUV missions.