Research On Key Technologies Of Digital Image Matting

Digital image matting has become a classic problem for digital image processing,which can be classified into two categories: blue screen matting and natural imagematting. In natural image matting, an initial mask called trimap is often required toroughly separate the input image into three regions: known foreground, knownbackground, and unknown region. However, due to the under-constrained problem andlimited extra hypotheses for image matting, the results of usual methods could alwayslead to much incorrectness compared with the ground truth matte, which could alsoheavily influence the final compositing results. Meanwhile, image matting also lacks ofa standard, effective and pixe-wise evaluation system. In this thesis, we mainly discusssome key techniques and their solutions, for the following issues: matting forsemi-transparent objects and environment light elimination in blue screen matting,pixel-wise evaluation for natural image matting, matting for a single natural image withdefocus background, and matting for the insufficiency of samples provided by trimap.In blue screen matting, due to these inaccuracy results in the solutions ofsemi-transparent objects and environment light interference for traditional methods, weraise a solution for semi-transparent objects based on blue distance compositionequation, and a solution for environment light elimination based on environment lightblending factor. For the former problem, the solutions of gray and color objects aregenerated according to the blue distance composition equation, respectively. For thelatter problem, we estimate the environment light factors and alpha values incorresponding regions, respectively. A new composite image is finally obtained througha new background. Experimental results show that our approach could generate moreprecise results in both the above two problems, for both alpha mattes and compositions,than previous methods.In natural image matting, due to the lack of a local evaluation scheme for alphamattes in the current image matting domain, we present an alpha evaluation systembased on four pixel-wise factors for these matting results. These evaluation factors arebased on the correlations of alpha results and original input images, and on boundarypriors of ideal alpha mattes. Then we raise a combination method of these optimizedalpha values from each pixel through a MRF framework. Experimental results show that,such an evaluation system could generate a pixel-wise judgement for the results of thestate-of-the-art matting methods without ground truth mattes. Moreover, a final optimalalpha matte could also be generated compared to the other alpha results.In natural image matting, due to the similarity between foreground and backgroundcolors in blurred backgrounds, we present a precise matting method for single natural image matting with a blurred background based on the matching of the secondderivative responses. In order for this, we firstly perform an edge detector to roughlypredict some edge pixels. Then for each edge pixel, we estimate its blur degree andjudge whether it belongs to a blurred background edge. If it does, we extend it along thegradient direction and mark those passed pixels as pure background pixels, orbackground blur priors. Finally, along with these extended background blur priors, wegenerate the final alpha matte through a general matting algorithm. Experimental resultsshow that our background blur priors could generate much more precise alpha resultsthan those state-of-the-art algorithms with the fact that those predicted backgroundpriors could become the key prevention of the similarity between some foreground andblurred background colors.In many cases, the unknown region always contains too much pixels for a providedtrimap, which could bring on insufficiency for foreground and background samples. Inorder for this, we make the basic trimap expansion, which marks some unknown pixelswith pure foreground or background, according to their color differences to the knownregion. A group of localized windows, which cover the entire unknown region andcoincide with each other, are initially generated based on the shape of the unknownregion. Then we study the expansion thresholds through learning and sampling areas forboth foreground and background regions in each window. The known foreground andbackground regions are thus extended respectively through their connectivity andsmoothness properties. Experimental results show that the basic expansion couldprovide effective samples for the matting process, and could lead to much more precisealpha mattes than previous matting methods.Finally, due to the insufficiency of known samples after the basic trimap expansion,we further make the advanced trimap expansion, which exploits some inner pureforeground and background pixels through themselves. In order for this, we raise twoconceptions of base-scene and sub-scene, which imply the simpler and more complexsides of foreground and background regions in each local window, respectively. Thesub-scene expansion is then decomposed hierarchically into two categories, which bothraise sample fitting processes through the unknown region themselves instead of knownregions, and are thus called self-learning. Experimental results show that the globalsub-scene pixels, generated from the extended expansion, could provide many extraefficient samples for matting, based on the basic expansion.