Reference-Based Multi-Stage Progressive Restoration for Multi-Degraded Images.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image restoration (IR) is challenging, especially with multiple distortions, due to the ill-posed nature of recovering high-quality details from degraded inputs.
• Existing deep learning methods struggle with complex, multi-stage degradations.

Purpose of the Study:

• To propose a novel multi-stage approach for progressive restoration of multi-degraded images.
• To introduce a Reference-based Image Restoration Transformer (Ref-IRT) that leverages texture transfer from reference images.

Main Methods:

• A cascaded U-Transformer network performs preliminary image recovery in a coarse-to-fine manner.
• Subsequent stages employ a quality-degradation-restoration method for accurate content/texture matching.
• A texture transfer/reconstruction network maps features from a reference image to enhance the target image.

Main Results:

• The Ref-IRT model demonstrates significant effectiveness in restoring multi-degraded images.
• Experimental results on benchmark datasets show superior performance compared to state-of-the-art methods.
• The proposed method successfully transfers relevant textures for enhanced image quality.

Conclusions:

• The Reference-based Image Restoration Transformer (Ref-IRT) offers a powerful solution for challenging image restoration tasks.
• Texture transfer from reference images significantly improves the restoration of images with multiple distortions.
• The approach provides a robust framework for advancing multi-degraded image restoration techniques.