Colorful Image Colorization with Classification and Asymmetric Feature Fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Automatic image colorization aims to convert grayscale images into colorful ones.
• Existing methods using regression loss yield undesirable brown tones, while classification loss can cause color overflow and requires extensive computation for color categories and balance weights.

Purpose of the Study:

• To propose a new method for computing color categories and balance weights for color images.
• To develop a U-Net-based automatic colorization network that addresses limitations of prior approaches.

Main Methods:

• A novel category conversion module and category balance module were developed to efficiently obtain color categories and balance weights, significantly reducing training time.
• A classification subnetwork was integrated to constrain the colorization network using category loss, enhancing colorization accuracy and saturation.
• An asymmetric feature fusion (AFF) module was introduced to merge multiscale features, effectively mitigating color overflow and improving overall colorization quality.

Main Results:

• The proposed U-Net-based colorization network achieved peak signal-to-noise ratio (PSNR) and structure similarity index measure (SSIM) of 25.8803 and 0.9368 on the ImageNet dataset.
• Experimental results demonstrate that the algorithm produces colorful images with vivid colors and higher saturation.
• The method effectively prevents significant color overflow, outperforming existing algorithms.

Conclusions:

• The proposed automatic colorization method offers a significant improvement over existing techniques.
• The novel modules and network architecture contribute to enhanced colorization accuracy, saturation, and visual quality.
• This work provides an effective solution for high-fidelity grayscale image colorization.