Multiscale and Multilevel Feature Fusion for Small Object Detection in Sonar Images

Target detection in sonar images holds significant implications for ecological conservation, marine engineering, and military applications, etc. However, existing object detection models which are mostly developed based on fixed receptive fields are challenged by detecting multi-scale objects, especially detecting small-scale objects from low-quality sonar images. To address this issue, we propose a novel multi-scale and multi-level feature fusion framework built upon You Only Look Once (YOLO), namely YOLO11s-MSML for object detection in sonar images. This architecture performs feature fusion from varying receptive fields within the same layer while enabling efficient cross-layer fusion of shallow and deep features, thereby strengthening detection capabilities across scales, particularly for small objects. Multi-scale spatial features are extracted using different-sized convolutional kernels via a Receptive Field Block (RFB) module within the same layer; multi-level features from cross-layers are integrated through a novel vertical module combining Content-Guided Attention (CGA) and SimAM attention with slicing operation (SWS). In addition, for model lightweight design, Distribution Shifting Convolution (DSC) is introduced to improve the model efficiency. Experimental results on the public Sonar Common Target Detection dataset (SCTD) demonstrate our model's superiority over existing algorithms, achieving 96.6%mAP0.5 for all objects. Compared with the state-of-the-art method, our algorithm improved detection accuracy of small objects by 3.3%.