Invertible Image Decolorization.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Invertible image decolorization is a valuable technique for color compression in multimedia systems.
• Current methods aim to generate faithful grayscale images that allow for complete restoration of the original color version.

Purpose of the Study:

• To propose a novel color compression method using invertible neural networks (INNs) for generating invertible grayscale images.
• To enable efficient separation and encoding of color information into Gaussian distributed latent variables.

Main Methods:

• Utilizing invertible neural networks (INNs) to separate color information from images and encode it into Gaussian latent variables.
• Incorporating wavelet transformation into a UNet-like INN architecture for effective grayscale learning.
• Implementing a quantization embedding to prevent information loss during format conversion and enhance generalizability.

Main Results:

• The proposed method successfully generates invertible grayscale images.
• Demonstrated state-of-the-art performance on three benchmark datasets in both qualitative and quantitative evaluations.
• Achieved efficient recovery of the original color version by re-sampling Gaussian variables and using the reverse mapping of INNs.

Conclusions:

• The novel INN-based approach offers a superior solution for invertible grayscale image generation.
• The method significantly improves efficiency and fidelity in multimedia communication and storage systems.
• The framework shows strong generalizability for real-world applications due to enhanced information preservation.