Object segmentation has been a challenging problem in computer vision and also plays an important role in many applications, such as image classification, object recognition, video surveillance, etc. In recent years, graph-theoretic segmentation methods have attracted much attention due to their solid theories and good performances. This thesis mainly emphasizes on object segmentation based on S/T Graph Cuts. We make some contributions on reducing user interactions and improving the practicability by introducing human vision related salient information into our framework. Besides that, we design a robust and effective attention extraction approach also based on S/T Graph Cuts. The contents and contributions of this thesis mainly include: 1.A comprehensive review for object segmentation and especially for segmentation based on S/T Graph Cuts is presented. The advantages and disadvantages of two typical S/T Graph Cuts methods are discussed through extensive experiments. 2.An automatic object segmentation framework, i.e. Saliency Cuts, is proposed for images with single salient objects. The proposed method integrates saliency detection and S/T Graph cuts which can supply efficient object and background labels and obtain complete and accurate segmentations automatically. 3.A "Local Saliency Cuts" framework is presented for segmentation of images with multi-salient regions or objects. Only a few interactions are involved to get satisfactory results, and the computing time is also reduced for the local manner. 4.An Asymmetrical Graph Cuts model is designed to extract the interesting attention regions for personalized image browsing in mobile devices. As a deformation and extend of object segmentation, the proposed model can supply robust and effective attention extraction compared to previous approaches.
修改评论