An underwater image enhancement method based on learnable text prompt guidance
By using a learnable text-guided approach, we can explicitly distinguish the degradation types of underwater images and align feature representations in the frequency domain. This resolves the optimization conflict in mixed training with multiple degradation types and improves the enhancement effect and naturalness of underwater images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN MARITIME UNIVERSITY
- Filing Date
- 2026-06-05
- Publication Date
- 2026-07-17
AI Technical Summary
Existing underwater image enhancement methods suffer from problems such as conflicting optimization objectives, insufficient differentiation between different degradation types, and poor naturalness of enhancement results when training with multiple degradations.
We employ a learning-based text-guided approach, constructing a degradation-aware labeled dataset, using image encoders and text encoders for collaborative training to determine the main degradation type, aligning feature representations in the frequency domain, and optimizing the network by combining multiple degradation losses.
Explicitly distinguishing different degradation types enhances the model's degradation recognition capability and enhancement specificity in complex underwater scenes, improves the detail clarity, color reproduction and visual naturalness of underwater images, and reduces local over-enhancement and structural distortion.
Smart Images
Figure CN122416232A_ABST