Semantic Contrast for Domain-Robust Underwater Image Quality Assessment.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Underwater image quality assessment (UIQA) faces challenges from complex degradations and domain shifts.
• Existing no-reference IQA methods often rely on subjective and costly Mean Opinion Scores (MOS), limiting their applicability.
• Generalizing IQA models to diverse aquatic environments without human annotations remains a significant hurdle.

Purpose of the Study:

• To propose SCUIA, an unsupervised UIQA framework that predicts image quality without human annotations.
• To leverage semantic contrastive learning for capturing implicit correlations between image degradation and quality.
• To enhance model generalization and domain adaptability for robust UIQA.

Main Methods:

• Introduced a vision-language contrastive learning strategy to align image and textual features in a unified semantic space.
• Developed a hierarchical contrastive learning mechanism combining statistical priors and semantic prompts for improved quality discrimination.
• Implemented an unsupervised domain adaptation module using local statistical features to guide CLIP fine-tuning for disentangling domain-invariant quality representations.

Main Results:

• SCUIA achieves significant improvements over existing UIQA methods on public benchmarks.
• The framework demonstrates superior generalization capabilities across unseen domains.
• Unsupervised domain adaptation enables effective zero-shot cross-domain quality prediction.

Conclusions:

• SCUIA offers a robust and unsupervised approach to underwater image quality assessment.
• The proposed semantic and hierarchical contrastive learning strategies effectively address UIQA challenges.
• The framework shows strong potential for real-world applications requiring reliable quality prediction in diverse underwater conditions.