Research On Natural Image Matting Guided By Visual Perceptual Characteristics

As an important technology of image processing, digital matting technology aims to extract foreground objects from static images or video clips by the form of alpha matte, in order to facilitate the operation of composing the foreground with a new scene, and get a new composite image or video. As a hot topic of image processing, matting is getting more attention from us. There are two reasons for this phenomenon. First, it is related to the integration of computer vision and machine learning and provides basis for image editing. It has gained increasing applications in industrial design and home entertainment, so it has wide range of applications and dramatic commercial value. Second, extracting foreground object is a very challenging part in graphics and computer vision research. Complex details of object shape, alpha channel, occlusion and illumination change and disturbance from similar background makes this work more complicated. With decades of study, researchers have proposed many matting algorithms. But how to achieve robust and accurate extraction of foreground object from complex environment still needs further study. And this makes matting has great theoretical research value. Focusing on the deficiencies of the existing matting technologies, this paper keeps research on some natural image matting algorithm which are guided by the characteristics of human visual perception and some detail problems.For applications like film production and family entertainment, we applied the research of human visual perception characteristics into matting algorithm, which is belonged to cognitive psychology and neurobiology field. From this perspective, we try to design natural image matting algorithm, in order to reduce user interaction as much as possible and at the same time, get high quality matte result. This paper makes research on the follow four aspects:1. Starting from the optimization of user input, we investigate the influence of visual characteristics to matting algorithm and study convenient interactive method to reduce the work of user;2. We take a research on the trimap refine method which is based on image structure characteristics, in order to reduce method’s dependence on user interaction. We improve the Bayesian matting to increase its processing ability to complex images;3. We explore the sampling method which is guided by global saliency. This method can increase the robust matting method’s accuracy;4. We propose a local matting method based on structure information, with which we can refine the matte locally to improve the final result. Different from the existing method, this paper emphasizes to analysis the potential information of foreground object by human perception. Our goal is reducing the dependence between algorithm and user input, and improve the method’s accuracy at the same time.Based on the above objectives, the main innovation points of this paper are as follows:First, we proposed a semiautomatic matting algorithm based on the human visual attention model, in order to dell with images with simple background and single target object. Foreground objects in the image is usually the interest object in the visual field. This paper presents a matting method which use focus of attention as priori. By improving the visual attention model, the shift of attention focus will be controlled in the range of the interest object. Using the focus of attention as the foreground image constraints can maximize the prospects of providing prior, while avoid the blindness of user interaction. Then we use closed form matting algorithm to get the high quality image matte. This algorithm is applicable to natural images with only one foreground object.Second, we proposed a matting algorithm using local structure features. The accuracy trimap is a precondition for matting to achieve good results. However, it will spend a lot of human effort to obtain an accurate trimap. This paper proposed a triamp refinement method, which can overcome the existing algorithms limitation of cannot modeling the structural information and reduce the requirements of accuracy trimap input. We refine the trimap by means of labeling several points of the unknown area to already known foreground and background points, in order to provide more sufficient samples for matting algorithm. The traditional Bayesian matting is a method based on statistical learning, it use the Gauss mixture model to present the sampling information of unknown region. Bayesian matting compute each point’s alpha value separately, and the result matte is discrete. This paper improves this method by adding a smoothing item into the maximum likelihood function, and the experimental result prove this improvement can eliminate the discontinuity of matte. This algorithm is suitable for the natural images that background with complex texture and difficult to artificial divide precise trimap.Third, we research on visual saliency based sampling method and confidence based matting algorithm. We propose an improved robust matting method, which is conbined the advantage of sampling-based and propagation-based method. It overcome the sampling process of sampling along the edge of region and focus on collecting salient point in order to ensure sample of integrity. Then we compute the confidence of sample point. We use linear hypothesis, color space distance and local saliency as the basis for evaluation, select the samples with high confidence value to estimate the matte. Finally, we use random walk algorithm to estimate alpha value. The experimental results show the effectiveness and accuracy of the proposed algorithm.Fourth, we research on semi-supervised learning based local matting method. We propose an improved learning based matting method. In the process of local image color-transparency model learning, joined the adaptive coefficient considering the image blocks color and texture features to adjust the relationship between the data term and smoothness term in the objective function. Because to the image characteristics itself, in the high contrast areas, data term in the objective function should play a more important role. This algorithm is applied to the existing matting method as local correction method, the experimental results show the effectiveness of this local matting method.