Abstract:To address the challenges of low image resolution, significant variation in target size, water turbidity, and occlusion that lead to poor detection accuracy in underwater object detection, a YOLO11n underwater object detection model based on detail enhancement and multi-scale feature fusion (DEMF-YOLO11n) is proposed. The model is built upon the YOLO11n architecture. To mitigate feature loss caused by standard downsampling in low-resolution images and small targets, the robust feature downsampling (RFD) method is adopted to replace the shallow and deep strided convolution layers. In addition, a module in the backbone is redesigned using the poly kernel inception module (PKIModule) and context anchor attention (CAA), enhancing feature extraction for occluded and multi-scale targets, as well as improving perception in complex underwater environments. To address the blurring of target edges caused by water turbidity and similar factors, shallow texture features are first refined using the edge enhancer module (EEM), and are subsequently fused with deep semantic features in the head network via the content guide attention fusion (CGAFusion) module. Experimental results on the RUOD dataset demonstrate that the proposed DEMF-YOLO11n achieves a 2.8% improvement in mAP₅₀ and a 4.1% improvement in mAP_50-95 compared to the baseline model, with an increase of only 0.53M in the number of parameters.