Automatic Image White Balance Technique

In research fields of image processing and computer vision, color is animportant kind of information. However, color is a kind of unstable image feature.When the environmental light condition changes, the apparent color of a samephysical surface in the scene changes, too. This phenomenon brings problems andchallenges to both the color image visibility improvement and high-level computervision applications which depend on the color features. In human vision system,different apparent colors of a same surface can be percepted as a relatively constantcolor. This property is called color constancy. With a similar function of colorconstancy, computers may be able to maintain or restore the true color of the imageduring image processing, or to obtain the color features of an image correctly andstably. The problem that color is vulnerable to light affect can be solved.Computational color constancy includes two aspects: one is looking for acolor-constant color descriptor, the other is correcting the image color. The latter oneis also called white balance. This dissertation aims to research color constancyalgorithms for image white-balance processing. A improved algorithm has beenproposed and an automatic image white balance prototype system has beenimplemented. The main works of this dissertation is as follows.(1) Research on classic illumination estimation algorithms. Mastering basicconcepts of color constancy theory and principles of illumination estimationalgorithms. In this dissertation, experiments with public image datasets asexperimental material are preceded to classical algorithms. Analyzing and comparingthe differences of theoretical basis and performance among basic color constancyalgorithms. Gray World, White Patch, Shades of Gray, and other basic algorithmsachieve better illumination chromaticity estimation performance under certainconditions which meet the assumptions of each algorithm, thus yield better colorcorrection result. Starting from those typical algorithms, their strengths andweaknesses will be analyzed.(2) An automatic white balance algorithm based on achromatic surfacedetection is proposed. On the basis of studying a series of single-light-sourcedillumination chromaticity estimation algorithms, the product channel method isapplied to more accurately detect the achromatic surfaces in the image. With the white pixels set, the illumination chromaticity is estimated; the weightedproduct-channel method and the iterative bilateral threshold method are used tofurther optimize the white pixels set, which solve the problem of the setting of thethreshold for white pixels set and the influence of highlight regions in the image.Experiments with public image datasets show the improvement of illuminationestimation accuracy of the proposed algorithm compared to classical algorithms.(3) Color constancy algorithms are applied to low-illumination videoprocessing. Low-illumination image or video processing is widely used in imageprocessing fields. This dissertation applies automatic white balance algorithms basedon color constancy into video processing. The illumination color of each video frameis estimated, then the video color correction is effectively achieved; Moving-averagetechnique is used to solve the blinking problem of the output video, and frames areadaptively enhanced in accordance with the brightness of original frame. The imagequality of low-illumination video is effectively improved. Experiments show thesystem also achieves real-time requirements.