The Research And Application Of Image Matting

Natural image matting approaches rely on user guidance and prior assumptions onimage statistics to establish the model for solving the values of alpha matte, and thenextract the foreground element from the background. How to establish a fast, accurate andeffective matting model is the focus of natural image matting research.In this thesis, we research the image matting technology, and analyze and comparethe pros and cons of several natural image matting algorithms. Our main contributions areas follows:Firstly, we category the image matting methods as sampling-based methods,propagation-based methods and methods that combine sampling and propagation, and listseveral typical methods and analyze.Secondly, we study and analyze KNN matting which is belonging to thesampling-based methods. This algorithm achieves a better balance of speed andeffectiveness, it is also able to obtain good matting results when the image has acomplicated background. But matting quality declines when the colors are a little differentor foreground and background layer is complexly mixed. In this paper we try to improvethe KNN matting by adding texture feature to enhance the foreground and backgroundconstraints in calculating the pixel feature vector, then the dimension of eigenvectorincreases from six to seven. Experiment results show that the matting quality is improved.Thirdly, we propose a method named an matting method based on color distance anddifferential distance which combines sampling-based method and propagation-basedmethod, then we gives the theory and compare it with several methods to illustrate thesuperiority. The experimental results have proved it.Finally, we apply the principles of closed-form matting in the colorization ofgrayscale images. Through the colorization experiments on various test images, it isconfirmed that the colorized images obtained by our proposed method have more vividcolors and sharper boundaries than those obtained by the traditional method.