Research on forward-looking sonar target detection algorithm based on edge enhancement and multi-scale feature fusion

Forward-looking sonar (FLS) images suffer from low resolution, severe noise interference, and limited discriminability of small targets, posing significant challenges for underwater object detection. To address these issues, this paper proposes ESBN-YOLO, a detection framework built upon YOLOv11 with three key improvements. First, an Efficient Multi-scale Bi-directional Feature Pyramid Network (EMBSFPN-SC) is designed, which leverages a three-round bi-directional fusion strategy coupled with a zero-parameter spatial channel enhancement mechanism to strengthen multi-scale feature representation. Second, an Edge Information Enhancement Module (EIEM) is introduced, adopting a dual-branch architecture that fuses Sobel-based explicit edge features with deep semantic features, thereby sharpening the model's sensitivity to object boundaries in noisy sonar imagery. Third, the Normalized Wasserstein Distance (NWD) loss is combined with CIoU as a weighted-sum bounding-box regression loss, replacing the original IoU-only formulation, yielding improved recall and localization accuracy for small targets. Experiments on the UATD dataset demonstrate that ESBN-YOLO achieves an mAP@0.5 of 88.67%, outperforming all compared methods. Additional evaluations on the MDD and FDD datasets further confirm the generalization capability and practical applicability of the proposed approach.