Research On Color Image Segmentation Algorithm Based On Clustering And Region Merging

Image segmentation is dividing image into some regions.It's based on features in the image,such as color,space and texture.We should ensure that these features in same region have consistency or similarity,and between different regions have obvious differences after segmentation.Although many studies have got some achievements on color image segmentation algorithm,no general segmentation method can be applied to all the color images.In view of this situation,the dissertation puts up a method applying in color image segmentation that is combining clustering and region merging realised in the Lab color space.About the selection of color space,the dissertation transforms the image's space from RGB to Lab.We make a research and summary on the common color spaces.We've found that the RGB space and CMYK space belong to the equipment oriented.This kind color space is not uniform.It's not suitable for image segmentation and analysis.The HSI space and Lab space belong to visual perception oriented.Their independence and uniformity are good.They're suitable for color image processing.But there is still existing nonlinear in the H component of HSI space.Therefore,we choose the Lab color space for image processing.About the choice of image segmentation method,the dissertation has made a research and experimental simulation on the current kinds of image segmentation methods and proposes a method based on improved K-means clustering and region merging.Before segmentation,we use median filter to smooth the image and to remove the noise.During the improved K-means clustering,we choose a fixed large K value and divide each channel in ascending order of Lab space into K parts.Then we choose the highest color value in each channel as initial clustering center,and we can get an initial segmentation result.In the process of region merging,considering the color,edge,area and adjacent relations,we use a function of measure the regional distance to merge regions,and use a changing rate of divergence in the color image to control the termination of region merging.At last,we can get the final result of the color image segmentation.We apply the algorithm to the Berkeley Segmentation Dataset 300,and make simulation.The result shows that the efficiency of the algorithm is no better than the traditional K-means,but the accuracy of segmentation results is much better than traditional K-means.