A Feature-Enhanced Multi-Scale Attention Network for Marine Debris Detection in Underwater Imagery

Underwater marine debris detection remains challenging because light attenuation, scattering, low contrast, blurred object boundaries, and complex seabed backgrounds degrade feature representation and hinder the reliable detection of small, sparsely distributed, and irregularly shaped debris. This study proposes the PNEB–MAAN–DUEM Network (PMDNet) to address these challenges. PMDNet is a compact YOLOv11n-based detection framework, and it improves underwater feature representation by coordinating feature enhancement, attention-guided multi-scale aggregation, and spatial feature reconstruction. The Parallel NARU Enhancement Block (PNEB) strengthens edge-aware feature representation through Edge-Gaussian Feature Enhancement, adaptive attention, and normalized residual transformation. The Multi-scale Attention Aggregation Network (MAAN) enhances contextual feature learning by integrating the proposed Noise-Robust Energy Attention (NREA) mechanism with parallel depthwise spatial processing for multi-scale feature aggregation. The Depthwise Upsampling Enhancement Module (DUEM) reconstructs spatial feature information through single-stage 2× upsampling and lightweight feature refinement. Experiments on a 15-class underwater debris dataset show that PMDNet gains 6.3 and 5.4 percentage points in mAP@0.5 and mAP@0.5:0.95, respectively, over YOLOv11n (means of three training runs), and that it requires 10.2 GFLOPs. The results indicate that integrating feature enhancement, multi-scale contextual aggregation, and spatial feature reconstruction in a coordinated way improved aggregate detection accuracy on the evaluated dataset. PMDNet therefore provides a compact detection framework with an explicit accuracy–computation trade-off, while its benefits under specific underwater degradations require further validation.