Study On Example-Based Color Processing

With the increasing popularity of digital cameras and the rapid growth of World Wide Web, people can easily get large numbers of digital images and videos. Simultaneously, people are always eager to edit their huge scale libraries of images and videos by themselves. Therefore, the research for editing techniques for digital image and video becomes a hotspot in the computer graphics and computer vision communities. Researchers also focus on color processing, and many research results in the color processing domain are published in important academic journals and conferences'proceedings.In this dissertation, we focus on the color processing techniques for digital images and videos. Firstly, we review the recent works of color processing including color transfer, color-to-gray conversion, colorization and color hamonization. Then, based on the survey, we propose the study on example-based color processing which especially emphasizes the techniques'functionality, performance and user-friendliness.Our research focuses on digital images and videos'color, we thus discuss the theories about color vision and color systems in the second chapter which is the important theoretical basis for our study on example-based color processing. Our research mainly includes as follows:? Firstly, we propose an example-based color transfer approach in correlated color space for color images which extends Reinhard et. al's algorithm to correlated color spaces. The proposed algorithm conveys exemplar images'color to target images directly in correlated color spaces. We analyzes the input images'pixel values at first by using principle component analysis approach and obtains the manipulating matrices including the translation, scaling and rotation matrices. After applying these manipulating matrices sequentially on the target image, we get the resultant image which comprises the target image's scene and exemplar image's color characteristics.? From the defition of color transfer, we propose a gradient-preserving color transfer algorithm. Considering the fidelity in terms of the gradient map in the target image and the color distribution of the examplar image, we formulate the processing of color transfer as an optimization problem and resolve it. Meanwhile, we present evaluation metrics for objectively assessing the performance of global example-based color transfer algorithms. Our experimental results validate our methods and high fidelity.? Also, we present a temporal color morphing scheme to simulate some natural scenes which is characterized with smooth color alteration and frequently appeared in many movies. Restricted by the great time span of these natural scenes, the traditional film-making methods usually do not produce the real evolving process of the scenes. Our algorithm takes the several minutes long initiative part of the simulated scene as input, and converts it to lαβcolor space. Then we can treat with the input data's three components separately because the correlation between channels in lαβspace is minimized, especially for natural scenes. For the purpose of the algorithm's performance, we quantize the lαβcolor coordinates. Then, the algorithm produces the resultant image sequences with smooth color alteration by histogram morping. Users can control the color morphing speed by customizing velocity curves.? From the view of machine learning, we propose a RBF neural networks-based temporal color morphing scheme. Firstly we pre-process the paires of pixel-blocks selected by users and obtain the training data for RBF networks. Then, we train a RBF neural network by these data. Finally, we embed the user-specified velocity curve into the RBF model to produce the resultant sequence. The experiments show the RBF-based temporal color morphing method is a more flexible and robustic one and has higher performance.? Based on the color of images, we study tongue manifestation recognition in Traditional Chinese Medicine. We proposed a fast roughness-augmented segmentation algorithm for tongue images and a supervised manifold learning based recognition method of tongue manifestation. In the last chapter, we conclude the dissertation together with some further research plans and enhancements of the proposed methods.