Research On Graph Cuts Methods Based On The Edge And Region Feature Modeling

Graph Cuts model is currently one of the most widely used models in image segmentation. It has important research value and can be applied to many different applications. Traditional segmentation algorithms based on Graph Cuts model include the edge and region feature of the images implicitly in the model. However, it is difficult to obtain ideal segmentation result with this implicit constraint. This dissertation explicitly treats the edge and region as feature constrains into tradition Graph Cuts model, and by this method it can improve the quality of image segmentation. The main works and innovations in this dissertation are listed in the following aspects:(1)In this dissertation, Canny algorithm with a variety of different thresholds is used to extract the edge feature in the image. At the same time, the region feature could be obtained using over-segmentation algorithm. In order to eliminate the edge feature of the region in the image, we obtain the fusion of the edge and region feature. In this way, we can correct the extracted edge feature.(2)In order to make full use of the edge and region feature in the image, this dissertation will treat the edge and region feature as a constraint and apply them explicitly into the traditional Graph Cuts model. First, this method construct a weighted graph based on pixels and integrate the edge feature in the traditional Graph Cuts model. Second, in order to simplify the structure of the weighted graph and accelerate the speed of segmentation, this dissertation adopts a superpixel approach to construct the weighted graph, using the edge and region feature extracted from superpixel approach as a priori constraints and apply the constraints into the traditional Graph Cuts model.(3)Unlike the modeling method of the traditional Graph Cuts segmentation algorithm based on Graph Cuts model, this article adopts a kernel density estimation approach to model the distribution of the region feature. Bhattacharyya measurement method is adopted in this dissertation to measure the similarity between regions and between the region and the distribution simultaneously. This unified metric ensures a consistent units for calculation, and it can reduce the number of iteration for segmentation.To verify the feasibility and effectiveness of the proposed method, a large number of color images are tested in the experiment. Through comparative analysis, we found that the proposed approach is more robust and can fully utilize the edge and region feature to guide segmentation. In addition, the proposed approach significantly reduces the number of iterations and thus reduce the processing time.