Research On Image Co-segmentation And Co-matting

Vision is an important perceptual skill which is important in human evolution. Digital media comes with technological advancements and more and more digital media come to our life. As a result, a large amount of digital data, such as images and videos, accumulate. The demands for data post-processing are ever changing. It is hoped to make human life mode fruitful and convenient by using computer technologies for investigating human visual experiences,this pursuit gives rise to a new branch of computer application, the computer vision. Computer vision is to extend human visual experience to the virtual world such that a computer senses and analyzes the environment, and eventually gives appropriate responses. This means that a computer can have visual abilities, like object tracking, scene comprehension and pattern recognition, just like what humans can do with their vision. Image binary segmentation and matting are core technologies in computer vision and this article is an overview of this domain.Image segmentation is a process and technique to divide a target image into several non-overlapping areas that have independent characteristics based on a set of criteria. As one of the image segmentation techniques, image binary segmentation separates the foreground from the background. Image binary segmentation can be roughly categorized into interaction-based and co-segmentation techniques. In interaction-based technique, a user has to provide a priori knowledge as a framework for the segmentation to work on. As this requires a lot of interactions with the user, the co-segmentation technique emerges as a result for reducing user’s burden, aiming to yield better foreground/background segmentation by using the strategy of maximizing the similarity of the foreground.Image matting is another foreground extraction technology. Compared with the segmentation technology, matting can extract more accurate foreground, matting is widely used in the film production and medical diagnosis field. In image processing, at the junction of complex foreground and background, it is difficult for image binary segmentation to split the foreground from the background. Therefore, matting brings in the alpha channel to solve this problem. Given an input image, matting estimates the opacity value a as well as foreground F and background B with a well-known linear composition equation. The matting problem is ill-posed. To solve this problem, state-of-the-art technologies usually come up with some restrictions such as trimap. Existing image matting algorithms can roughly be classified to three categories, sampling-based approaches, propagation-based approaches and hybrid-model-based approaches. Sampling-based approaches assume each unknown pixel is blended by a color sample pair picked from F and B. Generally a sampling confidence measurement is involved to identify the best sample candidates, which are then utilized for estimating each alpha value individually by using the composition equation; Propagation-based approaches establish associations between alpha values of neighboring pixels by assuming foreground and background colors satisfy constant or linear model in local neighborhood. Alpha values of all unknown pixels are then jointly evaluated by known matte propagation from F and B without explicitly deriving foreground and background colors; Hybrid-model-based approaches first estimate an initial matte with a sampling based approach, followed by affinity evaluation between neighboring pixels with one of the propagation-based approaches. The two terms are then integrated in a unified Gibbs objective energy function-an optimization frame, from which, the final alpha matte is derived.In this paper, we propose an example-based semi-automatic image collection segmentation framework. Each target image is first automatically cutout by learning from the segmentation information of samples and followed by rectification process with user assistant if necessary. We validate the proposed framework by testing it on multiple standard segmentation datasets, which shows that it achieves better performance, comparing with co-segmentation techniques, and accurate object cutout with surprisingly few user interactions.State-of-the-art matting methods mainly focus on extracting a matte from a single image, inspired by recent advances in image co-segmentation, we propose a new framework called co-matting. Co-matting aims at simultaneously extracting high quality matte in a global optimization scheme from multiple images with similar foreground against dissimilar background, the target of the optimization is that the initial mattes produced by the single image matting algorithm are automatically improved with the aid of their counterparts in other images which have higher confidence. Experimental results show that this co-matting framework can achieve noticeably higher quality results on an image stack than applying state-of-the-art single image matting techniques individually on each image.