Image Segmentation Algorithm Based On Clustering

Image segmentation plays a very important role in the field of image engineering.Since its birth,image segmentation has attracted a large number of scholars to study and research.At present,a large numbers of different image segmentation algorithms have been proposed.However,existing segment ation algorithms are mostly designed for specific research objects,general theory of the division has not been procided yet,and the exploration of the new segmentation algorithm and segmentation theory still have a long way to go.There are two difficulties in the field of image segmentation based on clustering algorithm.First,the particularity and complexity of the image itself which makes the image segmentation algorithm has a lot of limitations.Usually,an algorithm only applicable to a certain kind of images.Second the size of pixels.Especially,with the rapid development of computer equipment,images have become more and more clear,the size of pixels become to increase,more and more details have been stored.People have higher and higher standards for the requirement of image segmentation in quality and speed.Many segmentation algorithms are time-consuming or have a low precision.This paper devoted to the research of clustering image segmentation algorithm,find new ways to save the above two problems.The innovation of the paper mainly presents in the following three aspects.1.We proposed an improved FCM image segmentation algorithm based on hierarchical clustering and peak detection.The FCM algorithm needs to determine the number of clust ers before segmentation.It is sensitive to noise and initial clustering centers,and it is easy to fall into locally optimal solution.The hierarchical clustering algorithm has strong adaptability and is not sensitive to the initial data set.New method d oes not require the number of cluster centers before clustering,and it can use the images histogram information to find appropriate cluster centers automatically.First,using images histogram generates a set of possible cluster center;then,decreases ev ery unbefitting center and generate new cluster centers via the iteration of peak detection and hierarchical clustering.After initialization,using FCM clustering partitions the image.Finally,experiment results prove that this algorithm is faster and mo re effective at image segmentation.2.We proposed an improved Nystr?m spectral clustering image segmentation algorithm based on error sampling.Spectral clustering algorithm has been a research hotspot in the field of image processing,recent years.Spectral clustering based on the similarity of data while structure of similarity matrix is complex.The calculation of spectral clustering can be very time-consuming,especially in the process of Eigen-decomposition for Laplacian matrix.Nystr?m extension method could obtain the approximation solution of eigenvectors by using a small amount of sample information,reduce the computational complexity of spectral clustering effectively.Based on the features of image and the error analysis of Nystr?m a new sampli ng method is presented.Using Uniform Sampling generates a set of cluster centers at first;then,minimize the error between data and centers by iteration;finally,experiments have been given with satisfactory result.3.We proposed an improved sparse matrix spectral clustering image segmentation algorithm based on multi-scales.There are two obvious problems in spectral clustering algorithm in image segmentation.First,the algorithm needs to calculate the similarity between the pixels and performs eigen-decomposition on Laplacian matrix which is computationally expensive and time-consuming.Second,the algorithm has a single distance scale when calculating the similarity matrix,and the result is poor in reliability.To solve these two problems,a new image segmentation algorithm based on multi scales sparse matrix spectral clustering is proposed.Based on the error analysis of similarity matrix,a new scheme of feature extraction is designed at first.Then extracted the feature of image in different distance scales,and the sparse similarity matrix is created by using the feature information.Finally,the accuracy and robustness of the algorithm are verified by theoretical analysis and image segmentation experiments.