Restoring Ambiguous Boundaries: An Efficient and Robust Framework for Underwater Camouflaged Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Underwater Camouflaged Object Detection (UCOD) faces challenges due to boundary ambiguity and optical degradation.
• Biological mimicry and aquatic environments obscure target edges, hindering detection accuracy.

Purpose of the Study:

• To develop a lightweight detector, CAR-YOLO (Camouflage Ambiguity Resolution YOLO), for improved UCOD.
• To effectively address boundary ambiguity and enhance object detection in complex underwater scenes.

Main Methods:

• Proposed a frequency-domain dual-path mechanism (FRM-DWT/EG-IWT) for high-frequency edge recovery.
• Introduced a Low-level Adaptive Fusion (LAF) module to integrate low-frequency semantic information.
• Implemented an Uncertainty Calibration Head (UCH) to refine supervision using prediction consistency.

Main Results:

• CAR-YOLO achieved 27.1% mAP50-95 on UCOD10K, outperforming SOTA methods with reduced parameters and GFLOPs.
• Attained 30.7% mAP50-95 on UWB-COT220, a 7.7-point improvement over YOLOv11.
• Demonstrated strong generalization capabilities through cross-domain experiments on UODD.

Conclusions:

• CAR-YOLO effectively mitigates boundary ambiguity in UCOD.
• The model achieves a superior balance of accuracy, robustness, and computational efficiency.