Research On Video Matting Algorithm

Matting refers to the technique of accurate foreground extraction in images and video. It is one of the key techniques in many image editing and film production applications. As the image matting is inherently an under-constrained problem, user guidance and prior assumptions on image statistics are generally needed to solve this problem. Some classical matting techniques proposed to capture the images/video against a single or multiple constant-colored background(s). However, these methods cannot be easily applied to the images which are captured in natural environment. In recent years, the image matting techniques in complex natural background are widely studied. As there is no restriction on the background, these techniques have more broad potential applications. In these methods, how to achieve accurate matting results with few user interactions within a relatively short time is still an open problem. Video matting can be regarded as a more complicated extension of image matting. However, accurate and fast extraction of dynamic objects from video sequences with few user interactions is still a challenging problem in the computer vision field.This thesis does some in-depth researches on the images/videos matting techniques. The main work and contributions are listed as follows:1) An improved blue screen matting algorithm with the joint bilateral filter is proposed based on the color difference keying method. This algorithm can achieve real-time performance by the parallel implementation with GPU acceleration. Experimental results demonstrate that this algorithm is fast and efficient.2) When the color line assumption is not satisfied, the guided image filter generally cannot obtain good matting results. This thesis presents three improved methods. The first method uses a nonlinear median filter on the smoothing results of the guided filter to make it better satisfy the color line assumption. The second method processes the images with two tandem guided image filters to improve the matting results when the original alpha input is not good enough. The third method, which is based on the color clustering ball model, extends the color line model within a small local window to the non-local feature space neighborhood. Experimental results demonstrate that these proposed methods can achieve good image matting results.3) Based on the Visual Background Extractor(Vi Be) algorithm and the improved guided filter, a novel automatic video matting method is proposed. Experimental results demonstrate that the proposed algorithm can achieve good results within a short computational time.