The Research Of Foreground Extracting Algorithm Based On Image And Video

In this thesis, we discuss about the technique about segmentation and matting of image and video. Segmentation and matting technique is widely used in industry, and are very applicative in computer application, it is widely used in film production and medical industry. For the film such as science fiction film and magic film like Harry Potter, it is very important to use the blue matting technique and video composition technique to achieve different special effects.Though blue matting technique is widely applied in industry, the usage of blue matting technique is limited because the blue background and special tools are all needed. Today, the matting technique in nature environment is studied extensively. This technique is hoped to produce perfect composition effect without special tools and special environment in the future.We introduce segmentation and matting technique in this thesis. Segmentation technique is to segment the image by binary value, foreground is represented by 0 , background is represented by 1, there are no transition area in the result of hard segmentation, so segmentation could not achieves high quality results sometimes. However, segmentation is always the prior procedure of matting, and it is importan-t. We introduce some important segmentation method based on Graphcut algorithm such as Grabcut and Lazysnapping, we implement the algorithms and enhance them to achieve better results. In GrabCut algorithm, we enhance the interactive method and achieve better results. In lazysnaping algorithm, we propose a new likelihood es- timation based on geodesic distance. Experiments show our method is effective. We also show the limitation of Graphcut algorithm which is the base of our segmentation method, many suggests are given in the thesis.We also discuss about the matting technique,we introduce some important al-pha matte technique such as blue matting and bayesian matting and poisson matting. There are two kinds of matting technique, one is based on sampling, the other is based on propagation. A good alpha matting algorithm should have low SAD and high s-moothness. In this thesis, We implement the closed form matting algorithm and robust matting algorithm which are the most effective matting algorithms in recent years, the first is based on propagation, the last one combine the sampling method and prop-agation method together. In our experiment, we show those methods can achieve high quality results by limited interactive input.We discuss the video segmentation and matting technique. Different from image segmentation and matting technique, video segmentation matting algorithm should maintain the temporal-spatial consistency. We introduce some important algorithms such as 3D graphcut algorithm, and implement the video matting technique by optical flow, the optical flow algorithm is used to track the trimap in each frame of the video, so the alpha matte is extracted on each frame of video by image matting technique. Benefited from temporal consistency, foreground and background pairs can be sam-pled from neighbor frames,so more accurate sampling can be found. Some example show our method is effective.Finally, we apply matting technique to nature scene, we propose a framework to extract the fog and rain and snow from video. At first, we find the features of those nature scene, then we can extract the object from scenes. We also reconstruct the background from nature scene. Examples show our method can achieve good results.We improve the old method by follows:(1) We enhance the interactive method in GRABCUT algorithm. Compared with old method, we get more perfect results.(2) We apply the geodesic distance in lazy snapping. Our likelihood function is more effective than old function.(3) We improve the sampling method in video matting technique, our method benefit from temporal consistency and achieve good results.(4) We propose a new method to detect the snow particle in the snow scene.(5) We propose a new framework to process the weather in nature scene, fog and rain and snow are all included.