Image clustering method and system

Abstract

The invention relates to an image clustering method, which comprises the following steps of: creating a directional graph for a provided image sample set by using a variable bandwidth non-parameter nuclear density evaluation; partitioning the created directional graph into at least two non-intersected sub graphs by using a random walking isoperimetric partition method; and extracting image data in the sub graphs, and classifying the image data in the sub graphs into one category. The image clustering method fully considers the local probability density information of image data distribution, and can effectively cluster the data distributed extremely non-uniformly; and because the non-parameter clustering method is used, the method can process the image data with irregular shape distribution. Moreover, the invention also relates to an image clustering system.

Description

【Technical field】 [0001] The invention relates to the field of pattern recognition, in particular to an image clustering method and system. 【Background technique】 [0002] Clustering refers to dividing a sample set without category marks into several subsets or categories according to certain criteria, so that similar samples can be classified into one category as much as possible, and dissimilar samples can be divided into different categories as much as possible. Cluster analysis is a kind of multivariate statistical analysis and an important branch of unsupervised pattern recognition. As an unsupervised classification method, cluster analysis has been widely used in many fields such as pattern recognition, data mining, computer vision and fuzzy control. Traditional clustering methods, such as K-means method and EM method (maximum expected value method) are all based on convex spherical sample space, but when the sample space is not convex, the method will fall into local...

AI Technical Summary

Problems solved by technology

Although the graph-based spectral clustering method has achieved some success, the Gaussian kernel-based spectral clustering method using a fixed bandwidth cannot obtain satisfactory clustering results on image sample sets of many natural scenes, even if the parameters are carefully adjusted. These methods cannot handle multi-scale sample sets well, and cannot effectively reflect the local probability density distribution of image data.

Images