Acoustics-YOLO: A Multiscale Feature Fusion-Based Method for Small-Target Sonar Image Detection in Shallow-Water Environments

Abstract

In near-shore shallow-water environments, continuous autonomous underwater vehicle (AUV) motion causes substantial acoustic target scale variations, while seabed sediments, rocks, and clutter severely degrade sonar image quality. These factors collectively increase the difficulty of underwater small-target detection and recognition. To address these challenges, we propose an underwater acoustic target detection framework based on multiscale feature fusion, termed Acoustics-YOLO. Specifically, this paper designs a cross-scale feature pyramid network (CSFPN) architecture that effectively exploits bidirectional interactions between low-level detail feature maps and high-level semantic feature maps. Through this cross-scale information exchange, more efficient aggregation of multiscale information is achieved, thereby improving the capability for small-target detection. Furthermore, to mitigate the parameter growth and computational overhead introduced by multiscale designs while effectively suppressing background noise during feature fusion, we propose a lightweight background-suppressed group shuffle convolution (BS-GSConv). Subsequently, this study collected 9819 forward-looking sonar images covering 4 target categories in real shallow-water environments and subsequently constructed a dedicated forward-looking sonar dataset. Experimental results demonstrate that the proposed model achieves an improvement of 0.8% in mAP@0.5 and 1.1% in mAP@0.5:0.95 compared with the baseline model. In particular, it improves mAP@0.5 by 2.4% for the small-target category Roundtable. For the Shoal category, which exhibits high acoustic similarity to background clutter, it achieves a 0.8% improvement in mAP@0.5. Delivering real-time detection performance at 243 frames per second (FPS), the model effectively balances detection accuracy and computational efficiency. Notably, the proposed network was further validated across different YOLO variants and open-source datasets, with the detection results confirming the effectiveness of the method. Therefore, the proposed method satisfies the requirements for real-time AUV detection and perception in near-shore shallow-water environments.