All-in-One Image Restoration via Causal-Deconfounding Wavelet-Disentangled Prompt Network.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Standard image restoration methods face challenges with high storage costs and known degradation patterns.
• All-in-one image restoration (AiOIR) models unify multiple degradations but suffer from spurious correlations and biased degradation estimation.

Purpose of the Study:

• To address the limitations of current AiOIR models.
• To propose a novel method for effective and generalizable all-in-one image restoration.

Main Methods:

• Introduced Causal-deconfounding Wavelet-disentangled Prompt Network (CWP-Net).
• Employed wavelet attention modules in the encoder and decoder to disentangle degradation and semantic features.
• Utilized a wavelet prompt block for causal deconfounding to correct biased degradation pattern estimation.

Main Results:

• CWP-Net effectively disentangles semantic and degradation features, mitigating spurious correlations.
• The proposed method corrects biased degradation pattern estimation through causal deconfounding.
• Experimental results demonstrate superior performance of CWP-Net over state-of-the-art AiOIR methods.

Conclusions:

• CWP-Net offers a robust solution for all-in-one image restoration.
• The causal deconfounding approach significantly enhances model effectiveness and generalization.
• This work advances the field of image restoration by addressing key limitations in current AiOIR models.