Research On Co-segmentation Method Based On Multi-sources Attributions Of Image

How to get the target quickly which user interested in from the more complicated images is becoming the hotspot and difficult point in the field of computer vision and pattern recognition. Image segmentation has always been the extensive attention of scholars as a way to quickly extract target, and it has made many research results. However, there are still many difficulties. For example, the image contains increasingly information, and the features of image are also progressively rich, one single feature cannot satisfy today's technical requirements. In terms of segmentation granularity, the segmentation methods based on pixel tend to split the complete target, and the segmentation methods based on region or super-pixel tend to loss the detail of target, which is heavily dependent on the accuracy of the pre-segmented region. To address these problems, the co-segmentation method based on multi-sources attributions of image is discussed in the dissertation. In details, the main innovative research achievements of this dissertation can be described as follows.Firstly, a new effective texture feature modeling method is proposed based on the nonlinear compact multi-scale structure tensor(which is proposed based on the nonlinear compact modeling of the traditional multi-scale structure tensor) and total variation flow. Then, the Grab Cut framework with the proposed texture descriptor is applied to deal with the texture image segmentation, and the corresponding experiment results with the traditional texture descriptors verify the superiority of our proposed texture descriptor in terms of high efficiency and accuracy, what is more, its lower dimensions compared to the multi-scale structure tensor can accelerate the subsequent probability modeling.Secondly, the co-segmentation frameworks for multi-sources attributions of image is proposed based on virtual node, and it is used to deal with the multi-feature co-segmentation problem. The color texture image co-segmentation with virtual node is implemented based on the L*a*b* color space and our proposed texture descriptor. And the corresponding experiment results with the energy mixture model or the single feature verify the superiority of our proposed co-segmentation method. The co-segmentation framework can absorb the advantages of the different feature segment results adaptively.Finally, the co-segmentation frameworks for multi-sources attributions of image is proposed based on content information, and it is used to deal with the multi-granularity co-segmentation problem. The multi-granularity co-segmentation with content information is implemented based on image pre-segment regions by the improved mean shift and the image pixels. And the corresponding experiment results with the single granularity verify the superiority of our proposed co-segmentation method in terms of details and target integrity.To verify the practicability and usability of the texture feature modeling method, the multi-feature co-segmentation method with virtual node, and the multi-granularity co-segmentation method with content information, a large number of simulation experiments have been presented in this dissertation, and these techniques have good application prospect.