Color and Frequency Correction for Image Colorization

TL;DR

This work analyzes DDColor and identifies two key limitations: uneven performance across frequency bands, especially in high-frequency regions, and color cast due to insufficient input dimensionality. It proposes two optimization schemes: (i) frequency-band-specific coloring with an artifact-removal post-processing module based on a 4-level SEB-enhanced U-Net, and (ii) a color-cast correction network that leverages mean-color inputs via an encoder–decoder bridge with an $L_1$ loss. Quantitatively, the approach yields improvements in PSNR and SSIM across frequency bands, with an average PSNR gain of about $+1.30$ dB and a notable high-frequency gain of approx. $+1.21$ dB; a concatenated model offers further gains at the cost of training time. Overall, the paper advances image colorization by improving high-frequency detail and color fidelity, while proposing user-friendly mean-color inputs to stabilize and potentially control the coloring process.

Abstract

The project has carried out the re-optimization of image coloring in accordance with the existing Autocolorization direction model DDColor. For the experiments on the existing weights of DDColor, we found that it has limitations in some frequency bands and the color cast problem caused by insufficient input dimension. We construct two optimization schemes and combine them, which achieves the performance improvement of indicators such as PSNR and SSIM of the images after DDColor.