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

Area of Science:

• Robotics
• Computer Vision
• Marine Technology

Background:

• Forward-looking sonar (FLS) images present challenges like low resolution and noise, hindering underwater object detection.
• Small targets in FLS imagery are difficult to distinguish due to limited discriminability.

Purpose of the Study:

• To develop an advanced detection framework for improving underwater object detection in FLS images.
• To enhance the accuracy and robustness of detecting small and noisy targets.

Main Methods:

• Proposed ESBN-YOLO framework based on YOLOv11.
• Introduced Efficient Multi-scale Bi-directional Feature Pyramid Network (EMBSFPN-SC) for multi-scale feature representation.
• Developed Edge Information Enhancement Module (EIEM) for improved boundary sensitivity.
• Utilized Normalized Wasserstein Distance (NWD) loss combined with CIoU for bounding-box regression.

Main Results:

• ESBN-YOLO achieved an mAP@0.5 of 88.67% on the UATD dataset, surpassing existing methods.
• Demonstrated superior performance in detecting small targets with improved recall and localization.
• Validated generalization capability on MDD and FDD datasets.

Conclusions:

• ESBN-YOLO effectively addresses the limitations of FLS imagery for underwater object detection.
• The proposed enhancements significantly boost detection performance, especially for small and noisy targets.
• The framework shows strong generalization and practical applicability for marine applications.