Research On Algorithms For The Interactive Image Segmentation And Semantic Image Segmentation

Image segmentation is a traditional problem of image processing and computer vision.The results of image segmentation directly impact on subsequent image processing. It canbe said that image segmentation is the key step from image processing to image analysis.In recent years, research have mainly focused on four types of image segmentationalgorithms, which are interactive image segmentation algorithm, particular class imagesegmentation algorithm, semantic image segmentation algorithm and co-segmentalgorithm. In this paper, we study interactive image segmentation and semantic imagesegmentation. Many problems in computer vision, including image segmentation, can benaturally phrased in terms of energy minimization. In the last few years researchers havedeveloped a powerful class of energy minimization methord based on Graph Cut.Firstly, almost all interactive image segmentation algorithms based on Graph Cuthave the problem of Short Cut. To solve the problem, we use Gaussian Mixture Model(GMM) and Shortest Geodesic Distance jointly to model the appearance of image. And weuse Geodesic Star Convex Shape as the priori object shape constraint. We combined theappearance model and the shape constraint in Conditional Random Fields, CRF, and thenwe construct the energy function of our interactive image segmentation algorithm whichcan be optimized by using Graph Cut.Secondly, to solve the problem of the closed universe recognition paradigm ofsemantic image segmentation in the category training phase, we choose the concept ofnonparameters in the category training phase. In our semantic image segmentationalgorithm, we just only need to calculate and statistic image features. When there are newimages or new categories to add the training image set, we simply calculate the newfeatures of image or statistic image features related to the new category.Finally, to solve the problem of large amount of calculation in semantic segmentingthe test iamge, we oversegment the test image, then we use suprepixels as the basic unit touse KNN algorithm to calculate the posterior probability of the superpixel belongs to aspecific categories. And then we use the joint probability of the category of the adjacent superpixel as the cotextual inference. We combined the posterior probability and thecotextual inference in Markov Random Fields, MRF. At last we construct the energyfunction of our semantic image segmentation algorithm which can be optimized by usingGraph Cut.