Abstract: In water emergency rescue tasks, the effective range of optical imaging devices is often severely limited or may even fail due to water turbidity and light attenuation. Therefore, forward-looking sonar (FLS) has become a key sensor for detecting underwater human targets. However, FLS images are affected by severe speckle noise and multipath effects, while annotated human acoustic data are extremely scarce, resulting in high missed-detection rates and poor generalization under small-sample conditions. targeting human body sonar imaging characteristic, a physics-guided human target detection algorithm based on an improved YOLO11n network is proposed for underwater forward-looking sonar images in small-sample scenarios. A Wavelet-based Downsample Block (WDB) is introduced into backbone network to suppress frequency-domain noise and preserve target contours. A Shadow-aware Holistic Acoustic Attention (SHAA) module is designed to jointly model highlight echoes and acoustic shadows, so that sonar imaging priors can be injected into deep features. A Dynamic Acoustic-Context Module (DACM) is further used to capture diverse human target shapes through dual-branch geometric modeling. Experiments on a self-collected underwater human dataset show that mAP50 is improved by up to 35.6% compared with YOLO11n, while the number of parameters is increased by only 3.1%. Missed and false detections are reduced while real-time performance is maintained, providing an effective approach for rapid underwater human target detection in water emergency rescue tasks.

Key words: underwater forward-looking sonar, object detection, sonar image, lightweight network, water rescue