Neural Degradation Representation Learning for All-in-One Image Restoration

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Current image restoration methods excel on single degradation types but fail when degradations are unknown or mixed.
• Real-world image restoration is challenged by unknown and varied degradation types, leading to significant performance degradation.

Purpose of the Study:

• To develop an all-in-one image restoration network capable of handling multiple, simultaneous image degradations.
• To address the limitations of single-degradation models in practical, uncontrolled environments.

Main Methods:

• Proposed a novel neural degradation representation (NDR) to capture and decompose diverse degradation characteristics.
• Developed a degradation query and injection module to effectively utilize learned degradation representations.
• Implemented a bidirectional optimization strategy to alternately train degradation and restoration processes.

Main Results:

• The proposed network demonstrates effective performance across multiple degradation types, including noise, haze, rain, and downsampling.
• The neural degradation representation (NDR) adaptively decomposes and addresses various image corruptions.
• Experimental results validate the method's effectiveness and generalizability for diverse image restoration tasks.

Conclusions:

• The developed all-in-one network with neural degradation representation (NDR) offers a robust solution for handling multiple image degradations.
• This approach overcomes the limitations of single-degradation models, improving performance in practical applications with unknown degradations.