Research On Image Segmentation Algorithm Based On Fuzzy Clustering And Mean Shift

Image segmentation technology is playing an increasingly important role in computer vision applications,various image segmentation algorithms are continually being improved and optimized to suit specific application scenarios,and the novel image segmentation algorithms have being proposed.Because of its simplicity and stability,fuzzy clustering algorithm has being widely used in image segmentation community.The robustness to various noises and the protection to image details are two important measurements for the performance of the image segmentation algorithm.In order to further improve the ability of suppressing noise and protecting image detail of the image segmentation algorithm,we propose the density constraint in chapter 3,which is on the basis of spatial-gray constraint.In chapter 4,we propose a membership degree smoothing method to enhance the consistency of the partitioned region.To use fuzzy clustering algorithm,two important constraints needs to be met,the clusters in the feature space of the data set should approximate to the spherical distribution,and the number of data clusters need to be set in advance.For the first constraint,the Euclidean distance as similarity measure of fuzzy clustering is replaced by Gaussian kernel distance in the third and fourth chapters of this paper,and the adaptive kernel bandwidth is used.The introduction of the kernel measure improves the robustness of the clustering segmentation algorithm to the pixel outliers and the noise,making the image partition more accurate.For the second constraint,the fourth chapter uses the mean shift algorithm to search the density peak in the image feature space.The density peaks are the cluster centers,mean shift algorithm can determines the number of cluster centers and clustering centers simultaneously,then we use the fuzzy clustering to segment the image,and achieve the unsupervised segmentation of natural images.This paper mainly has the following two works:1.In order to further improve the performance of suppressing the noise of the algorithm and to better protect the image details,this work has made three improvements based on the existing spatial clustering algorithm based on spatial and gray constraints.The first point of improvement is to propose the idea of density constraints.The second improvement is to use the Gaussian kernel instead of the European distance.The third point of improvement is to use the gray level instead of the gray value to participate in the iterative calculation.These three improvements improve the robustness of the algorithm to various types of noise,while better preserving the details of the image and reducing the amount of computation.2.In order to realize the unsupervised segmentation of natural images,a segmentation method incorporating mean shift and fuzzy clustering is proposed in chapter 4.The mean shift algorithm searches density peak iteratively in the direction of the density gradient in the pixel feature space,and finally determine the number of clusters and the optimal clustering center.Then fuzzy clustering algorithm is used to segment the image,which does not need to iterate the cluster center,just calculate the membership of the pixel,reducing the computational complexity of fuzzy clustering.Since the search starting point of the mean shift has a great influence on the search results,a method of initialize the initial point uniformly is proposed.In order to improve the consistency of the partitioned region,a membership degree smoothing method is proposed.The above two algorithms do not need to manually set the control parameters,as a result,their versatility is relatively strong.