Assessment of Automatic Hair Colorization and Relighting Using Chromaticity Distribution Matching

Abstract

This paper presents a new approach to human hair colorization and relighting. Human hair colorization, concerning given model hair image without changing neither hairstyle nor hair texture, is challenging. The fundamental problem making this task complicated is the difference in the hair texture and the illumination between a user and model images. Natural human hair consists of a mix of hair swatches. Each swatch has its chromaticity distribution, which, generally, is non-Gaussian. The proposed method treats these swatches as color clusters in the hair image. In this case, matching the user and the model hair swatches or color clusters solves the problem. After this matching, the color transfer between the relevant model and user swatches is applied. Besides, the model’s hair should be compressed to a reasonable size to provide simultaneous representation for numerous hair colors. The model’s hair colors are taken from the images of hair color packs that usually are available in decorative cosmetic stores. These images, however, are taken in standard illumination conditions, so appropriate relighting should be applied to provide a photorealistic user’s appearance. Experimental results with 530 different color models and more than 20,000 users show that the proposed technique achieves high photorealistic perception and a reasonable compression ratio. On average, a high pick signal to noise ratio (39 dB) indicates just a noticeable difference between original and reproduced model hair color.