Research On The Method Of Region-based Color Image Segmentation

Image segmentation is one of the critical steps from image processing to imageanalysis, which is also the primary problem in the process of image analysis andimage recognition. Color plays a vital role in the field of computer vision andpattern recognition, since color is an important feature for recognizing images.More and more favor and attention of researchers is paid to color imagesegmentation technology with the cost reduction of sampling color images, theimprovement of computer processing ability and the increase of applications.Segmentation technology for gray images has been mature and the segmentationtechnology for color images can be regarded as the extension of the segmentationtechnology for gray images, namely, the common image segmentation technologiesfor gray images are applied to the color space and the image segmentationtechnology is generated.The traditional clustering algorithm using Gaussian Mixture Model (GMM)has been used widely in the filed of image processing. However, the imagesegmentation for real color images is sensitive to the Gaussian noise and isolatedpoint since the spatial correlation between pixels is ignored. A segmentationalgorithm for color images using GMM clustering algorithm based on region patchis proposed in the dissertation and the segmentation accuracy is improved.The main work in the dissertation is as follows:(1) The definition and process of image segmentation is introduced and themain image segmentation algorithms are researched. The common color space andthe transformation methods between different color spaces are researched, and thecolor spaces introduced here are compared.(2) A segmentation algorithm for color images using GMM clusteringalgorithm based on region patch is proposed. First, the color gradient is computedfor color images. Second, the watershed segmentation algorithm is applied to thegradient images and some over-segmentation homogeneous regions are generated.At last, the region patch features are extracted which are treated as the inputsamples for GMM clustering algorithm and the segmentation can be realized. A newinitialization method using EM algorithm which imitates the feedback principle inthe controlling theory is used in the process of estimating the GMM model parameters, namely, the final parameters estimation is guided by initial estimationresults and clustering accuracy is improved efficiently.The clustering algorithms based on pixels are replaced by these using regionpatch features in the proposed segmentation algorithm and the influence for colorimages generated by noise and isolated points is reduced. Some tests are conductedusing some synthetic images and a number of real natural color images in the imagelibrary of Berkeley, relative to the classical segmentation algorithms, the testsresults demonstrate that the segmentation validity is improved efficiently using theproposed method.