Semi-supervised Clustering Based On Constraints For Images Segmentation

Image segmentation plays a vital role in image analysis and pattern recognitionbecause it is the major factor in determining the analysis results. For better imageanalysis results, the emerging semi-supervised learning has recently become a hot topicin machine learning with its capability to exploit and utilize prior information toimprove the efficiency and accuracy of image segmentation. Semi-supervised learningcan gain better learning performance by making full use of labeled samples andunlabeled samples to guide the learning process. Generally speaking, semi-supervisedlearning can be roughly divided into two categories, namely, semi-supervisedclassification and semi-supervised clustering. Semi-supervised clustering utilizes asmall number of labeled samples to guide the clustering process of unlabeled samples.We study the application of semi-supervised clustering algorithm in image segmentationin this paper. The contributions of this paper are threefold:(1) We propose a semi-supervised kernel K-means clustering algorithm based onpairwise constraints for image segmentation. Firstly, we utilize must-link constraints toinitialize clustering centers and it can keep the algorithm from falling into localoptimum. Secondly, the clustering process is extended to the kernel space. The datafrom the original input space are mapped to a high-dimensional feature space tocalculate the distance. As a result, the similarity of the data points in the same clusterand the dissimilarity of the data points in different clusters are increased. Meanwhile,the nonlinear boundary clusters can also be found in the original input space.Consequently, our algorithm can receive the result of image segmentation more quicklyand accurately. Experimental results have shown that our segmentation algorithm canget better results compared to the conventional COP-Kmeans algorithm andConstraintSelector algorithm.The visual effects have also been improved.（2）We have also proposed a scheme for image segmentation using WeightedSemi-Supervised Kernel FCM(WSSKFCM). The new approach is inspired bysemi-supervised clustering approach using the kernel-based method based on KFCM,namely, the Semi-Supervised Kernel Fuzzy C-mean algorithm (SSKFCM). In theinitialization step, we utilize the variance between samples to initialize the kernelparameters, thus avoiding inappropriate parameters by the manual adjustment toinfluence the performance of the algorithm. Meantime, we use distributing density sizeof the data points as weighted value to gain better results for heaps-like or data sets oflarge discrepancy of every class specimen number. This significantly improves the accuracy of image segmentation. Finally, we have verified that the algorithm is superiorto three other algorithms on synthetic texture images and simulated SAR images.(3) An image segmentation algorithm based on semi-supervised reformulatedfuzzy local information C-means clustering (SSRFLICM) is presented, which makesbetter use of the characteristic of RFLICM(Reformulated FLICM) together withsemi-supervised clustering. In the initial steps, the cluster centers are initialized byusing labeled samples. Meanwhile, we separately initialize the membership degreematrix of labeled samples by utilizing the hardening method and randomly initial themembership degree matrix of unlabeled samples to refrain from local optimum. The testresults on texture images, simulated SAR images and natural images have demonstratedthe efficacy of the proposed method.This paper was supported by the National Natural Science Foundation ofChina(No.61303032), the Fundamental Research Funds for the CentralUniversities(No.K5051302065) and the Fundamental Research Grant for XidianUniversity（BDY121427）.