Turning Diffusion-Based Image Colorization Into Efficient Color Compression.

Area of Science:

• Computer Vision
• Image Processing
• Data Compression

Background:

• Stroke-based colorization methods reconstruct images from sparse data, suggesting compression potential.
• Diffusion-based compression inpainting missing regions, closely related to colorization but underexplored.
• Existing colorization techniques are often integrated into transform-based codecs like JPEG, not diffusion methods.

Purpose of the Study:

• To bridge the gap between colorization techniques and diffusion-based image compression.
• To develop a novel luma-guided diffusion framework for anisotropic colorization in the YCbCr color space.
• To propose a new luma preference codec for efficient image compression.

Main Methods:

• Demonstrated the link between discrete colorization and continuous diffusion-based inpainting in YCbCr color space.
• Developed a luma-guided diffusion framework to steer inpainting in chroma channels based on luma structure.
• Proposed a luma preference codec allocating more bits to the luma channel for high-quality reconstruction.

Main Results:

• Anisotropic luma-guided colorization significantly outperformed Levin et al.'s method.
• The proposed luma preference codec achieved high-quality color reconstruction.
• The new codec surpassed state-of-the-art diffusion-based compression and rivaled transform-based codecs.

Conclusions:

• Luma-guided diffusion inpainting offers superior anisotropic colorization.
• A luma preference codec effectively leverages human visual system sensitivities for compression.
• The developed method presents a competitive alternative for diffusion-based image compression.