Latent Denoising Diffusion Models for Colorization of Historical Portraits Photographs

In the early stages of photographic technology, due to technical limitations, most photographic images were presented in grayscale.Early color photographs, when exposed to harsh environmental conditions over time, would gradually degrade, leading to the permanent loss of valuable image content.In recent years, with the development of computer-based digital imaging and scanning technologies, it has become possible to restore the colors of historical photographs.However, the color restoration of historical photos requires professional expertise and is a labor-intensive, expensive, and time-consuming task.Therefore, developing automated network models for old photo colorization and restoration holds significant importance.The latest Denoising Diffusion Probabilistic Models (DDPMs), based on probabilistic modeling, analyze the diffusion relationships between pixels in an image while preserving the image's details and structure.These models have shown remarkable performance in various emerging fields of computer vision.This paper proposes an latent denoising diffusion probabilistic network-based automatic colorization method by studying deep learning-based colorization techniques and DDPMs.A VQVAE-based latent variable model is introduced into the forward diffusion process and reverse inference of the DDPM, serving as a transitional bridge during the reconstruction process.This model is capable of converting black-and-white images into color images.Experimental results demonstrate that the latent denoising diffusion probabilistic model achieves optimal colorization performance on facial image datasets.Finally, this colorization technique is extended to the colorization of real historical portraits