Twin Adversarial Contrastive Learning for Underwater Image Enhancement and Beyond

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Underwater images exhibit significant distortion, negatively impacting object detection accuracy.
• Current enhancement methods often prioritize visual restoration over task-specific performance, sometimes degrading detection effectiveness.

Purpose of the Study:

• To develop an underwater image enhancement method that is both visually appealing and optimized for object detection tasks.
• To address the limitations of existing methods that fail to improve detection accuracy despite visual enhancement.

Main Methods:

• Proposed an object-guided twin adversarial contrastive learning framework for underwater image enhancement.
• Developed a bilateral constrained closed-loop adversarial enhancement module for unsupervised learning and feature preservation.
• Integrated a task-aware feedback module using detector gradient information to guide enhancement for improved detection.

Main Results:

• The proposed method significantly enhances visual quality of underwater images.
• Object detection accuracy using various detectors improved notably on images enhanced by this method.
• Demonstrated improved performance in high-level tasks like semantic segmentation through object guidance.

Conclusions:

• The object-guided twin adversarial contrastive learning approach effectively enhances underwater images for both visual appeal and downstream detection tasks.
• Task-aware guidance is crucial for bridging the gap between general image enhancement and task-specific requirements.
• The method offers a promising solution for improving underwater computer vision applications.