Learning Dynamic Prompts for All-in-One Image Restoration.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• All-in-one image restoration aims to unify multiple degradation types in a single model.
• Existing deep learning models struggle with heterogeneous degradations due to static prompts that fail to capture unique input attributes.
• Current methods often produce suboptimal results by averaging degradation characteristics.

Purpose of the Study:

• To develop a novel dynamic prompt approach for all-in-one image restoration.
• To overcome the limitations of static prompts in accurately representing diverse degradation priors.
• To enhance the performance and adaptability of unified image restoration models.

Main Methods:

• Propose Degradation Prototype Assignment and Prompt Distribution Learning (DPPD).
• DPPD decouples degradation prior extraction into Degradation Prototype Assignment (DPA) and Prompt Distribution Learning (PDL).
• DPA uses predefined prototypes for discriminative representations; PDL models prompts as distributions for adaptive sampling.

Main Results:

• DPPD framework achieves significant performance improvements across various image restoration tasks.
• The dynamic prompt approach demonstrates superior ability in capturing unique degradation attributes compared to static methods.
• Experimental results validate the effectiveness of the proposed DPA and PDL components.

Conclusions:

• The DPPD framework offers a more effective solution for all-in-one image restoration.
• Dynamic prompt learning is crucial for handling heterogeneous degradations adaptively.
• The proposed method advances the state-of-the-art in unified image restoration.