Research On Fast Graph Cuts Algorithm Integrating Pre-segmentation Information

Image segmentation often refers to the process of extracting the interest region from background according to the simple or complex features of the image. Computer vision algorithm has been widely applied to various fields. The image segmentation technology has an important influence on computer vision field, and the accuracy of segmentation results directly affect the high level processes, such as image understanding.Graph Cuts optimization method solve the image segmentation problem by minimize the energy function, and it has many excellent features, such as a good ability to integrate different image features and excellent time efficiency, these advantages has received much attention from many researchers, since graph cut model has been widely used in image segmentation. However, in recent years, with the higher resolution image that mobile devices can capture, the image segmentation complexity reaches a higher level, and the segmentation based on graph cuts often has several drawbacks: 1) require more user interaction to ensure the accuracy of the foreground/background model; 2) the graph cuts' “shrinking bias” features often lead to poor segmentation result when segmenting objects with complex details; 3) high-resolution image make the graph contain too many redundant nodes and increase the time consumption. To address those problems, this thesis study the pre-segmentation methods, using deformed multiresolution graph cuts and super pixels to improve the segmentation accuracy respectively.Firstly, an interactive image segmentation method based on deformed multiresolution graph cuts with high accuracy and low time consumption is developed. The method regards the "shrinking bias" issue of traditional graph cuts as a benefit and makes full use of it by using the deformed multiresolution technique, which can also provide a partial solution to it incidentally. The input image is first coarsened deformedly to some low resolutions with the different width-length ratios simultaneously, and then Grab Cut method is applied on them to obtain the different segmentations. To sum up the differences of these coarse labeling results, a "weighted map" is constructed to present possibilities of each area for foreground or background, which can describe the object in details with high accuracy. Finally, the "weighed map" is used to refine the trimap for building the more accurate Gaussian mixture models and graph cuts model to assign the final segmentation labeling.Secondly, we also studied the super pixel as a pre-segmentation method in image segmentation. To conquer the instability of super pixel detection in some special images, we use saliency, gaussian statistics and other features to describe a super pixel. Based on these features, we designed a specific energy function to get a better segment results, then we use weighted map strategy to implement the segmentation method integrating different size super pixel, which can improve the accuracy of the segment result significantly.Finally, in this thesis, our methods is evaluated on two famous benchmarks extensively. The experimental results indicate that our proposed methods has the higher segmentation accuracy as well as the lower time consumption when compared with the Grab Cut and even the recently proposed One Cut.