Color Correction by Estimation of Dominant Chromaticity in Multi-Scaled Retinex

Abstract

In image capture a scene with nonuniform illumination has an influence on the image quality, especially the contrast and detail in dark regions. Generally, the tone curve or histogram of an image is modified to improve the contrast and detail, yet this is insufficient as the intensity and chromaticity of the illumination vary with geometric position. Thus, the multi-scaled retinex algorithm has been proposed, where the influence of nonuniform illumination is reduced by partitioning the original image using local average images that are estimated based on Gaussian filtering of the original image. However, the multi-scaled retinex algorithm produces color distortion as the local average images are independently estimated for each channel. In particular, if the chromatic distribution of the original image is not uniform and is dominated by a certain chromaticity, the local average image includes not only the intensity and chromaticity of the illumination but also the dominant chromaticity through the Gaussian filtering, thereby distorting the color. Accordingly, this article proposes a multi-scaled retinex using a modified local average image to reduce the color distortion by the dominant chromaticity of the original image. As with the multi-scaled retinex algorithm, the local average image is obtained through Gaussian filtering of the original image. The local average image is then divided by the average chromaticity value of the original image to reduce the influence of the dominant chromaticity. However, because the average chromaticity value includes the dominant chromaticity of the original image and the chromaticity of the illumination, the chromaticity removed from the illumination in the local average image needs to be compensated. Therefore, the chromaticity of the illumination is estimated based on the chromaticity of the highlight regions in the original image. The chromaticity of the local average image is then modified by the estimated chromaticity. In experiments, the proposed method was found to improve local contrast and reduce the color distortion.