UWLight-YOLO: An adaptive lightweight attention network for underwater sonar image object detection

Underwater sonar image object detection is a key enabling technique for marine exploration, fisheries monitoring, and underwater robotic navigation. However, sonar imagery is often affected by severe noise interference, low contrast, blurred boundaries, and complex background clutter, leading to incomplete target representation and making scale variation difficult to characterize, which in turn limits the feature extraction and multi-scale modeling capability of conventional detectors. Meanwhile, resource-constrained underwater platforms impose strict requirements on model size and computational cost, calling for a lightweight detection solution that better balances detection accuracy and computational complexity. To this end, this paper proposes UWLight-YOLO, a lightweight object detector based on YOLOv8 for underwater sonar image analysis, aiming to reduce model complexity while maintaining competitive detection accuracy. Specifically, we first propose two lightweight C2f variants, C2f-Fast and C2f-FAMA, which replace the repeatedly stacked Bottleneck units in the original C2f with PConv-based FastBlock and EMA-recalibrated FAMABlock, respectively, thereby substantially reducing redundant computation without changing the input/output resolutions or channel configurations and improving feature representation robustness under complex sonar degradations. Subsequently, we propose the Coordinate-Enhanced Screening Feature Pyramid Network (CSFPN), which enables efficient multi-scale feature screening and fusion in the neck, thereby enhancing robustness under low signal-to-noise ratio conditions and blurred-boundary scenarios. In addition, we propose a Task Adaptive Dynamic Detection Head (TADDH) to strengthen information interaction between classification and localization, further reducing parameters and computation while improving the consistency of classification and regression predictions. Extensive experiments on the UATD dataset demonstrate that, compared with YOLOv8n, UWLight-YOLO reduces parameters by 61.3% and GFLOPs by 50.6% while maintaining comparable detection performance, validating its effectiveness and practicality as a deployment-oriented lightweight solution for underwater sonar object detection.