FEFM-YOLO11: Underwater Object Detection Algorithm Based on Improved YOLO11

• Published in: Neural processing letters Vol. 57; no. 5; p. 82
• Main Authors: Wang, Qi, Liu, Zhichuan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.10.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Complex Systems; Computational Intelligence; Computer Science; Design; Environmental monitoring; Feature extraction; Localization; Modules; Object recognition; Occlusion; Performance evaluation; Semantics; Target detection; Telematics; Underwater; Water quality; Loss function; CAFM; YOLO11; Object detection; FEM
• ISSN: 1573-773X; 1370-4621; 1573-773X
• DOI: 10.1007/s11063-025-11805-2

Underwater object detection technology is extensively applied in subsea resource exploration and benthic environmental monitoring; however, the underwater environment presents significant challenges due to limited visibility caused by insufficient illumination and turbid water quality, severely impeding detection tasks. To address these challenges and enhance precision, this paper proposes FEFM-YOLO11—an enhanced YOLO11-based underwater object detection algorithm. Primarily, the algorithm substitutes two C3K2 modules in the backbone with a Feature Enhancement Module (FEM). Employing a multi-branch dilated convolutional structure, the FEM increases feature diversity, expands the network’s local receptive field, and enhances semantic representation of small objects, thereby strengthening detection capabilities. Subsequently, a Context-Aware Fusion Module (CAFM) is introduced in the Neck, which effectively models global and local features through integrated local feature capture (convolutional operations) and global feature extraction (attention mechanisms), consequently improving denoising performance. Furthermore, a dedicated small-target detection layer is incorporated to specifically boost detection performance for smaller underwater objects. Finally, the Wise IoU loss function was used for comprehensive evaluation and performance optimization, and the collaborative integration of these components effectively reduced the problem of missed detections caused by occlusion in dense cluster targets. Experimental results on the URPC2020 dataset demonstrate that the improved algorithm achieves enhancements of 2.0% points in mAP@50 and 1.9% points in mAP@50:95 compared to the baseline, confirming that FEFM-YOLO11 elevates detection precision and validates the feasibility and effectiveness of the proposed methodology.