An Improved K-Means Clustering Method For Color Image Segmentation

Image segmentation is the key stept from image processing to the image analysis. The aim is to put the image segmentation image segmentation for multiple overlapping and each with different characteristics and the region, mainly used in image compression target extraction pattern recognition, etc. Previous image segmentation technology is mainly used in gray image, along with the rapid development of computer technology, color image gain more and more simple, applied more and more. Color image segmentation gradually concerned.Based on previous color image segmentation problem of a lot of the basis of research achievements, This paper proposes a will K-MEANS clustering algorithm、ant colony algorithm and the combination of watershed image segmentation method. The method of ant colony algorithm first of global searching capability and robustness and combined with the advantages of edge information, to determine the clustering center and the cluster number. Then use the watershed of the algorithm based on gradient of the original color images for the pre segmentation, and take the original color image data into some color consistency with the subset; The last of these subsets of center for K-MEANS clustering. The experimental results show that:Because subset of the original image is far less than the number of pixels, The clustering sample number significantly less, Improve the clustering speed, Can be effective sound of color image segmentation. At the same time in the clustering of characteristic distance to replace the Euclidean distance, improve the algorithm robustness. The main work the following:1. With characteristic distance as to clustering sample difference of the measure.2. Will color images from RGB color space transformation to the HSI color space.3. Using Of ant colony algorithm acquire initial clustering number and initial clustering center.4. Based on the gradient of the watershed of the initial segmented results proceed K-MEANS clustering.Compare with the traditional K-MEANS clustering algorithm, The proposed method is effective to overcome the cluster number must be based on prior knowledge set in advance of the initial clustering center is randomly selected cluster effect quality depends on the defects of the judge distance formula.