Research On Illumination Chromaticity Estimation In Computer Vision

Illumination changing will cause object's surface color changing and shadow in images and video clips. This will lead incorrect results for many computer vision tasks, such as image segmentation, object recognition, target tracking and so on. So illumination color estimation and adjusting is an important research area. This dissertation focuses on illumination chromaticity estimation problem in computer vision. Our work is carried out from the following three aspects.At present, illumination chromaticity estimation algorithms could be divided into two classes, unsupervised methods and supervised methods. Among unsupervised methods, Gray Edge using higher-order image structure performs better than the other algorithms using zero-order image structure. SVR-based illumination estimation algorithm produces the best results among supervised methods. So in this dissertation, we proposed a new illumination chromaticity estimation algorithm which combines Gray Edge and SVR together. This algorithm is proved to be able to improve accuracy of illumination chromaticity estimation.Almost all of the illumination chromaticity estimation algorithms are designed for still images. Although they could be used on video clips frame by frame, the relative information between frames is abandoned. So an illumination chromaticity estimation algorithm especially for videos is proposed. This method takes advantages of the similarities of illuminations and contents between adjacent frames. Experiments showed that using the proposed method generate better illumination chromaticity estimation results on video clips than the other methods.Shadow is a kind of multi-illumination situation. Illumination in shadow region is totally different with the illumination in non-shadow region. So shadow removal process is a multi-illumination chromaticity estimation and adjusting process. A shadow removal algorithm based on regional illumination chromaticity estimation is proposed here. Pixel values in shadow region are transformed to what they should appear under non-shadow illumination. Experiments on synthetic images and real images proved the proposed algorithm is effective.