Self-Supervised Colorization Towards Monochrome-Color Camera Systems Using Cycle CNN.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Monochrome cameras offer superior image quality compared to color cameras.
• Colorization aims to enhance monochrome images using reference color images.
• Existing methods often rely on simulated data, leading to performance gaps on real data.

Purpose of the Study:

• To develop a self-supervised model for accurate colorization of monochrome images using real data.
• To overcome limitations of training on synthesized data in monochrome-color camera systems.
• To improve the quality of color images obtained from monochrome cameras.

Main Methods:

• Introduced Cycle CNN, a self-supervised Convolutional Neural Network (CNN) model.
• Employed a dual colorization process using the Weighted Average Colorization (WAC) network.
• Incorporated cycle consistency loss, Global Curve Adjustment (GCA) network, structure similarity loss, and spatial smoothness loss for training on real data.

Main Results:

• Cycle CNN effectively trains directly on real data from monochrome-color camera systems.
• The model leverages cycle consistency to ensure accurate color channel reconstruction.
• Enhanced structural similarity and spatial smoothness were achieved for the colorized images.

Conclusions:

• The proposed Cycle CNN significantly outperforms existing methods for colorizing real-world monochrome images.
• Self-supervised learning on real data enables robust colorization without ground-truth information.
• This approach offers a practical solution for obtaining high-quality color images from monochrome sensors.