Domain Adaptation for Underwater Image Enhancement

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Learning-based algorithms excel at underwater image enhancement but struggle with domain gaps.
• Models trained on synthetic data often fail to generalize to real-world underwater images due to inter-domain and intra-domain gaps.
• Existing techniques produce artifacts and color distortions on real underwater images.

Purpose of the Study:

• To propose a novel Two-phase Underwater Domain Adaptation network (TUDA) to address both inter-domain and intra-domain gaps.
• To enhance the generalization capability of underwater image enhancement models for real-world scenarios.
• To reduce visually unpleasing artifacts and color distortions in enhanced underwater images.

Main Methods:

• TUDA employs a two-phase approach: triple-alignment for inter-domain adaptation and easy-hard adaptation for intra-domain gap reduction.
• Phase one utilizes adversarial learning for image, feature, and output-level adaptation, bridging the synthetic-real data gap.
• Phase two incorporates a rank-based quality assessment and pseudo-labeling for easy-hard adaptation, minimizing real data distribution gaps.

Main Results:

• TUDA significantly minimizes both inter-domain and intra-domain gaps in underwater image enhancement.
• The proposed rank-based quality assessment accurately evaluates enhanced image perceptual quality.
• Extensive experiments confirm TUDA's superiority over existing methods in visual quality and quantitative metrics.

Conclusions:

• TUDA offers a robust solution for real-world underwater image enhancement by effectively addressing domain adaptation challenges.
• The network demonstrates improved generalization and reduced artifacts compared to current state-of-the-art techniques.
• TUDA advances the field by tackling the often-overlooked intra-domain gap in underwater image processing.