Toward Better Than Pseudo-Reference in Underwater Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing underwater image enhancement (UIE) methods often rely on paired datasets with imperfect pseudo-reference labels.
• This limitation hinders the performance of current UIE networks, creating a need for improved training strategies.

Purpose of the Study:

• To develop advanced loss functions for UIE that overcome the performance bottleneck caused by imperfect pseudo-reference labels.
• To enhance UIE network performance without resorting to more complex network architectures.

Main Methods:

• A plug-and-play hybrid Performance SurPassing Loss (PSPL) was formulated, combining Quality Score Comparison Loss (QSCL) and Depth-aware Unpaired Contrastive Loss (DUCL).
• QSCL guides enhancement by comparing image and region-level quality scores against pseudo-references.
• DUCL addresses severely degraded distant regions by leveraging scene depth information, pulling distant regions towards nearby high-quality regions and pushing them away from low-quality distant regions.

Main Results:

• The proposed PSPL significantly improved UIE performance compared to state-of-the-art methods.
• The effectiveness of PSPL was demonstrated even when used with a simple and lightweight UIE network architecture.
• Experimental results validated the superiority of the proposed loss function in enhancing degraded underwater images.

Conclusions:

• The developed PSPL offers a novel and effective approach to underwater image enhancement by focusing on advanced loss functions.
• This method successfully breaks through the performance limitations associated with imperfect training data in UIE.
• The PSPL framework provides a robust solution for improving the visual quality of underwater images, applicable even with simple network designs.