Study On The Graph Cuts Of Energy Minimization Based Precise Object Segmentation For Image And Video

Foreground segmentation for image and video has been increasingly emerging in the field of computer vision. As a key technique, image and video segmentation can supply the most basic features for those applications based on image and video analysis, such as detection, tracking, recognition, etc. And also it plays an important role in these applications. It is of great significance for us to fully study it. The main prior hypothesis for statistic based image and video segmentation are similarity and incoherence. Accordingly, the cues can be used in image segmentation are color, texture, contrast, etc. In addition, the cues can be used in video segmentation are motion, stereo, etc.Recently, a lot of researches have suggested that the method of modeling foreground and background with statistic methods, and making decision under the rule of Maximum A Posterior provided an effective approach to precise object segmentation for image and video. There are two major difficulties for statistic based segmentation. First, both the multimode background and foreground model, and those prior hypothesis are very difficult to describe efficiently. Second, how to make use of hypothesis and cues effectively still remains as an open question. The MAP-MRF method provides the precise segmentation for image and video with a significant theoretic solution. Under the framework, the contextual constraints can be sufficiently used for object segmentation. The graph model based energy minimum approaches can provide powerful tool for energy functions which describe the segmentation. By reviewing and analyzing those existing segmentation algorithms, this dissertation proposes a series of novel precise segmentation algorithms for image and video on the basis of the Graph Cuts.Specifically, the main contributions of this dissertation are as follows:(1) After thoroughly reviewing the Markov random field and Maximum A Posterior method, and the graph cuts based energy minimum approach, the graphic processing unit based accelerated graph cut algorithm is discussed. (2) By analyzing the requirement of object detection in infrared image, a accurate coarse-to-fine infrared object segmentation algorithm is proposed. First, the constraint object region is detected. Second, the graph cuts based accurate object segmentation is executed. Finally, the algorithm is applied in the hard measure system for metal material.(3) An interactive object segmentation algorithm using hierarchical graph cuts for color image is proposed. The method, which consists of coarse object extraction at highlevel and boundary portrayal at lowlevel, can provide the user with friendly user interface, and fit the requirement of both accuracy and efficiency. The user can give their interaction on different level and scale. Finally, the method is applied in the computer aided tongue diagnosis system for traditional Chinese medicine.(4) By analyzing the characteristics of infrared imagery under the condition of urban outdoor, the foreground and background models are constructed, which can overcome the challenge of low SNR, great change of polarity, and the existence of aperture around the object.(5) An adaptive foreground model, which consists of the spatial coherence and temporal persistence, is proposed to solve the problem of the similar color distribution between local background and motive foreground.(6) A conditional random field based model for video segmentation is proposed, which effectively fuses local temporal-spatial constraint of both observation and label data. The experimental results on several challenging scene prove the robustness and efficiency of the method.Concluding remarks are drawn, and the prospects of the research are given at the end of the dissertation.