Image Segmentation Based On MRF Regularization And Semi-Supervised Graph Kernel Clustering

Image segmentation refers to divides images into some different and non-overlapping regions with similarity properties of image color,gray value,texture,shape and feature set.It is one of the fundamental and core problem in computer vision.However,the complication of image itself brings ill-posedness to the segmentation problem.Normalized cut(NCut in short)as the representative graph clustering and Kernel K-means(KKM in short)as the representative kernel method,have been successfully applied to the problem.Markov Random Field theory(MRF in short)is one of the most commonly regularization methods in image data.In this paper,an improved scheme based on MRF and its optimization algorithm is proposed for two newly proposed graph/kernel clustering methods.First of all,the semi-supervised normalized cut based on the Hidden Markov Random Field(HMRF-SNCut in short)is applied to image segmentation.The contrast sensitive Potts regularization term is introduced to enhance the boundary,and a new model based on MRF regularization is proposed.In order to optimize the model,we construct an upper bound function for HMRF-SNCut and perform linear approximation,which can be solved by Graph cuts technique.The experimental result show that,compared to the original semi-supervised NCut,our new model has a significant improvement in the error rate and other statistical indicators.Then,on the basis of KKM,a potential function similar to Potts model is introduced to express the semi-supervised information.A new segmentation model is proposed by combing the semi-supervised KKM(SKKM in short)and smooth regulator.The new model is iteratively optimized based on graph cutting technique by using the idea of bounds optimization.The experimental results show that compared with the original SKKM,the segmentation results are improved in terms of statistical indexes such as error rate and contour matching.To sum up,this paper proposes two image segmentation models based on semi-supervised graph or kernel clustering,and gives an optimization algorithm based on Graph cuts technology.Our work not only provides new methods for image segmentation,but also proposes a new optimization algorithms based on Graph cuts technology for semi-supervised graph and kernel clustering.