Research And Application On Digital Matting

Digital matting is the key technology in image processing, non-linear videoediting and film-making applications. It refers to the problem of extracting theforeground objects in the images or videos accurately, which is an important issue inthe field of computer vision. According to whether there is restrictions on the shootingenvironment, matting technology can be divided into blue screen matting and naturalimage matting.Blue screen matting technology originated in the TV film production, foregroundobjects in the shooting, the background color used in the background of a known colorblue green. On behalf of the algorithm are: color-difference technique, chroma keyingmethod and triangulation matting.Natural image matting technology requirements are very low for the shootingbackground, applies to any natural background is a research hotspot in recent years. Itcan be roughly divided into color sampling-based matting method and pixel similaritypropagation based matting methods depending on how to use the image statisticalcharacteristics and the correlation of the nearest neighbor pixels. All the abovemethods share two major drawbacks: when foreground and background regions havesimilar or overlapped color distributions, the estimated α using known samplescollected from such regions is not reliable; In highly textured regions, color similaritymay not be able to represent the affinity of neighboring pixels reliably since strongedges in these regions may cause the affinities to decrease, thus blocking thepropagation of alpha.The research work of this thesis mainly in three aspects:1. Research and implement chroma keying algorithm for the edge problem of thealgorithm, some improvements are proposed. Reference to color-difference technique,replacing the color channel, combined with technologies such as dilation and erosionhave solved the problem successfully.2. For chroma keying needs much computation when applications to videomatting, relying on non-linear editing project, thereby taking full advantage of the parallel nature and high-speed floats operations of GPU hardware, to accelerate thealgorithm in the non-linear editing system.3. To solve the problem that traditional nature image matting methods do notperform well in highly textured regions, a texture clustering matting (TCM) algorithmbased on KNN matting is presented. Color and position information are leveraged byKNN matting method to acquire the affinity among neighboring pixels, which do notperform well in highly textured regions. TCM overcomes the problem by consideringtexture as a feature. The matting Laplacian is constructed by clustering neighbors inthe feature space which contains color, position and texture information. Finally, usingthe closed form solution solves the experiments on benchmark datasets indicate thatthe algorithm can extract the foreground object effectively in highly textured regions.