PMF-YOLO: a parameter-efficient multi-scale detector for underwater optical imagery under scattering and nonuniform...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Underwater optical images are degraded by absorption, scattering, and uneven illumination, leading to reduced contrast, blur, and weakened target boundaries.
• Existing detectors struggle with cross-scale feature imbalance in underwater environments, where high-resolution details are lost and low-resolution semantics are affected by haze.

Purpose of the Study:

• To develop a parameter-efficient multi-scale detector, PMF-YOLO, specifically designed to overcome the challenges of underwater object detection.
• To improve the accuracy and robustness of small-target detection in degraded underwater imagery.

Main Methods:

• PMF-YOLO integrates a Parameter-efficient Multi-scale Feature Fusion Feature Pyramid Network (PMFFPN) with channel control and a small-object head.
• Incorporates Cross-layer Channel and Saliency Attention (CCSA) for recalibrating boundaries and saliency.
• Utilizes an Optimized Depthwise Separable-wise Convolution Block (ODSSBlock) for efficient global modeling of spatially correlated degradation fields.

Main Results:

• PMF-YOLO achieved 78.1% mAP@0.5 on URPC2020 Test-A, 72.1% on URPC2020 Test-B, and 86.3% on DUO.
• Demonstrated significant improvements over YOLOv8n, exceeding it by 7.9%, 6.1%, and 3.2% respectively.
• Achieved superior performance with substantially fewer parameters compared to existing stronger detectors.

Conclusions:

• PMF-YOLO effectively addresses cross-scale imbalance and degradation issues in underwater feature pyramids.
• The proposed model offers a parameter-efficient and high-performance solution for underwater object detection, particularly for small targets.
• PMF-YOLO represents a significant advancement in enhancing the visibility and detection accuracy of objects in challenging underwater environments.