Study And Design On Texture Image Segmentation Algorithm Based On Clustering Analysis

Image is frequently and effectively used in people's daily communication.With the development of emerging technology such as Big Data and cloud computation,the generated and obtained image data in daily life is becoming increasingly large.Effective classification and management has been extremely urgent.Image segmentation is one of the keypoints in image processing,and the image segmentation based on texture features is becoming more and more important in the field of image processing in recent years.This kind image segmentation mainly includes two aspects,such as texture features extraction and feature classification.This thesis analyzes the image segmentation based on texture features from these two aspects:(1)Texture features extraction is the precondition of image segmentation.Because of the complex and various texture patterns,the results would not be ideal if we only tend to some single texture features extraction method.In this thesis,the fusion of the three methods,such as Gray Level Concurrence Matrix,Gabor wavelet transform and Local Binary Pattern forms fusion feature vectors,each of characteristic is complementary to each other and taking advantages of each method.Texture features extraction method of multi features fusion is validated in experiment,and it is moderately universal.(2)In the research of image segmentation based on affinity propagation clustering algorithm,the hardest problem is that the number of image data samples is huge,and using Affinity propagation clustering algorithm will cause memory overflow because it needs to construct a complete similarity matrix.In order to solve this problem,an integrated clustering algorithm is proposed in this thesis.The peoposed method includes three layers.In the first layer,the algorithm starts from the overall view,it uses sparse affinity propagation clustering algorithm for rough dividing and selects a number of representative samples which could reflect the structure information of the dataset.In the second layer,the algorithm uses the DBSCAN clustering algorithm to divide the representative samples in a small scale.In the third one,the main work is to integrate and merge for final integrated results of original data samples.The experiments prove that the integrated clustering algorithm could avoid the memory overflow in the process of large datasets by the proposed algorithm,and the integrated clustering algorithm integrate the advantages of the both algorithm which could automatically determine the clustering center and be used to process various data samples.Also integrated clustering algorithm could achieve a good result in texture images segmentation.(3)The fuzzy c-means clustering algorithm is easy to get a reasonable result in the face of nonlinear and fuzzy problems in image segmentation.However,there are also some shortcomings of FCM clustering algorithm,such as the impact of the dispersion degree of data samples and the "noise" data samples on the clustering results are ignored.So a fuzzy c-mean clustering algorithm based on relative density(Relative Density,RD_FCM)is proposed in this thesis.It completely considers the local information contained in the data samples and the relative tightness of each two adjacent samples.In addition,the effect of isolated data sample points is excluded.The experiments prove that compared with FCM clustering algorithm,RD_FCM clustering algorithm could improve the clustering effects and the rate of convergence.