Color Constancy using Image Regions

Color constancy is important for various applications such as image segmentation, object recognition and image retrieval where object color features are extracted invariant to the illumination conditions. Different color constancy methods have been proposed. These methods, in general, compute color constancy based on all image colors. However, not all pixels contain relevant information for color constancy. Eventually, biased pixel values may decrease the performance of color constancy methods. To this end, in this paper, we propose a method based on low-level image features using subsets of pixels. Hence, instead of using the entire pixel set for estimating the illuminant, only relevant pixels in the image are used. Therefore, prior segmentation is performed to learn for different image categories (e.g. open country, street, indoor) which pixel set (i.e. image parts) is most appropriate for a reliable estimation. Based on large scale experiments on real-world scenes, it can be derived that for certain categories, like open country and street, the estimation is far more accurate using image parts than when using the entire image.