MAS-YOLOv11: An Improved Underwater Object Detection Algorithm Based on YOLOv11.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Underwater target detection faces challenges like poor visibility, occlusion, and scale variation.
• Existing models struggle with complex backgrounds and overlapping targets.

Purpose of the Study:

• To develop an improved model for robust and accurate underwater target detection.
• To enhance performance in challenging aquatic environments with complex interference.

Main Methods:

• Introduced the C2PSA_MSDA module for enhanced multi-scale feature representation using dilated attention.
• Implemented an adaptive spatial feature fusion detection head (ASFFHead) for robust multi-scale object detection.
• Developed a Slide Loss function with dynamic sample weighting to improve hard sample learning.

Main Results:

• Achieved significant performance gains on the DUO dataset: 3.7% recall increase, 3% F1-score elevation.
• Attained mAP@50 of 77.4% and mAP@50-95 of 55.1% on DUO, outperforming the baseline.
• Demonstrated cross-domain generalization with 76% mAP@50 on the RUOD dataset.

Conclusions:

• MAS-YOLOv11 offers substantial improvements for underwater target detection tasks.
• The proposed enhancements effectively address challenges like occlusion and scale variation.
• The model shows strong generalization capabilities across different underwater datasets.