Ho-Hyoung Choi, Byoung-Ju Yun
In color constancy research, one of the well-studied problems is how to estimate and remove scene illumination chromaticity from image data. Throughout several decades, illuminant estimation methods have been developed by many researchers and institutions as evidenced in literature. These methods have yet to overcome several color distortion problems such as color leakage, luminance shift, hue shift and so forth. Accordingly, this article proposes an illuminant estimation method that combines preprocessing and CNN to achieve color constancy. The preprocessing uses an average image to normalize a given input image as a method of addressing both varying illumination and uneven background from a real-life scene image. A 3-layer CNN architecture is then used to estimate scene illumination efficiently. Given image patches as input data, the CNN works in a spatial domain, which does not require taking hand crafted features usually adopted for conventional methods. To create a more effective model for estimating scene illumination, the proposed method integrates feature learning and regression to institute one optimization process within the network structure. The experimental results prove that the proposed method delivers better performance of predicting scene illumination for color constancy over the conventional methods.
color constancy, estimating scene illumination chromaticity, CNN architecture, feature learning, feature regression, optimization process, network structure
Cite this paper
Ho-Hyoung Choi, Byoung-Ju Yun. (2019) Illumination Estimation for Color Constancy Using Convolutional Neural Network (CNN). International Journal of Signal Processing, 4, 6-8