An Affinity Based Matting Algorithm And The Research Of The Application

The matting problem is to extract the foreground image layer of the input image with the corresponding opacity information. Using this technique people can easily extract the foreground layer of the input image, and integrate it into any background scene to generate a synthetic image of high quality. With the development of science and technology, people are not satisfied with the2D display device on market because of the lack of stereo visual effect, and looking forward to more vivid display effect. In natural3DTV systems, the one adopting2D video with depth information is a promising technical approach at present. The key problem of the approach is how to generate the depth image sequence corresponding to the input2D video with high accuracy. Therefore, the extraction and optimization of the depth information is of important practical significance for natural3DTV systems.Foreign scholars have proved that matting technology can be applied to3DTV systems to improve the overall display performance. This paper summarizes existing mainstream matting algorithms and proposes an improved matting algorithm based on sampling and affinity, which proves to be effective to generate alpha matte output of high accuracy. And the author applies the kernel of the proposed matting algorithm into the depth image refinement problems. The main achievements of this paper include:(1)A new grouping scheme of spectral matting components based on Bayesian decision criteria. Calculate the matting components of the input image according to the spectral matting algorithm, and calculate the foreground and background probability density distribution function based on them. Then calculate the posterior probability of each matting component, and use the Bayesian decision criterion to group the matting components.(2)A new matting algorithm based on the combination of sampling and affinity, which applies the idea of algorithms based on sampling into the affinity based energy function framework. The author redesigns the cost function and the feature vector, and achieves better matting performance.(3)Propose to apply the matting algorithm kernel to the post-processing and optimization of depth images, and propose the idea of Quadmap creatively. According to the basic principles of stereo matching, divide the input depth image into different regions, and assign the error region caused by occlusion specifically. The algorithm achieves good results on testing sequences captured by the real-time system.