Color Image Object Segmentation Based On Graph

In this paper, we propose two kinds object segmentation based on graph:one is Interactive Segmentation Based on Super-pixels and Multi-cues (ISBS&M), which is used to segment single image interactively, and the other one is Object Segmentation Based on Conditional Random Field and Sparse Coding (OSBC&S) to do great image dataset object segmentation.ISBS&M mainly improves segmentation results from three contributions. First, we improve the computation process of edge costs, in which the contribution of adjacent nodes is incorporated. Traditional methods computing edge cost only use the raw data of its two terminals. Second, we propose an edge descriptor named edge continuity, which comes from the observation on the gradient magnitude relationship among the neighboring superpixels. Third, we combine color information, edge information and texture information into the graph cut framework effectively with learnt parameters. The result provides better boundary placement and stronger region connectivity. This leads to less user interaction needed to produce a desired segmentation.OSBC&S incorporates features using the bag-of-features and sparse coding techniques, bases on CRF, and combines multi-cues. Besides multi-cues in a similar way with ISBS&M, it improves algorithm from two contributions:First, Instead of directly operating at the pixel or superpixel, we advocate the use of patches and split patches as the basic processing unit. Patch features are effectively represented using spare coding based on the learnt pixel-level dictionary and spatial pyramid matching, while image patch contains more local structure characteristics to distinguish its categories than the superpixels which have be used in many current works, and split patches represent better adherence to object boundaries than the superpixels due to the latter uniform size constraints and homogeneousness. Second, we incorporate the unary classification probability as the unary inputs in a conditional random field, to refine the segmentations further using regions based on patch. High segmentation accuracy is demonstrated on three different databases:the Weizmann horse dataset, the VOC2006 cow dataset and the MSRC multiclass dataset, it shows that our approach achieves favorable results compared to the state-of-the-art approaches.