Research On Interactive Graph Theory Segmentation Method Based On Multi-scale Super-pixel Perception

Image segmentation often refers to the process of extracting the target objects of interest from the complex image environment according to certain similarity criterion and prior knowledge.It has been a key step in the research for image analysis,computer vision and other fields.In recent years,the interactive image segmentation method based on graph cut has attracted wide attention due to its simple interaction,accurate segmentation result,and has become one of the most important methods for the existing image segmentation.However,these methods generally construct segmentation model based on the neighboring relationships between pixels.Due to the lack of the structure information,it is difficult for them to correctly segment images containing texture and noise.In addition,the conventional graph cut based methods are sensitive to the seeds marked by user.When the seed's information is limited,it is hard to get satisfactory segmentation.To address the above problems,we introduce the image multi-scale information and superpixel-based higher-order information into the graph cut model.By combining with the likelihood diffusion and perceptual learning,we optimize the energy functions in the segmentation framework based on graph cut.The main work of this thesis includes the following parts:(1)To overcome the defects that the traditional pixel-based methods are sensitive to image texture,noise and seeds,we introduce the multi-scale region information into the conventional graph cut model,and propose an interactive segmentation algorithm based on multi-scale superpixel and graph cut.The structural information of superpixel is obtained by extracting the mean and covariance of all pixels in the corresponding region,so as to measure the relationship between the different superpixels more accurately.By fusing multi-scale superpixel information,the local neighborhood constraint of small-scale superpixel is utilized to overcome the over-segmentation,and the long range connectivity constraint of large-scale superpixel is utilized to overcome the under-segmentation.(2)To further improve the quality of the segmentation and the robustness to the seeds,we fuse the relationships among the pixel and the multi-scale superpixel,and propose an interactive image segmentation method based on likelihood diffusion and perceptual learning.The Gaussian mixture model is used to estimate the initial probabilities of pixels and superpixels belonging to the labels.In order to make full use of the limited seed information,the global similarity relationships of the image is captured by diffusing the neighboring similarity relationships between pairwise pixels,and a probability diffusion strategy is proposed to estimate label prior probabilities more accurately.Furthermore,the pixel-layer and the superpixel-layer information of the image are fused in the segmentation framework to maintain the local neighborhood and spatial continuity of the segmentation results.