Research Of Adaptive Image Segmentation Based On Scale-space Theory

Image segmentation has always been a fundamental step in image processing, it helps to bring great convenience to the more advanced image analysis, making the intelligent, speedy, accurate image analysis possible. As a widespread used tool, it has gained great value in intelligent transportation, military research, biomedical processing, modern agriculture, automation and so on. Image segmentation has always been a hot research topic in recent years.Among all the image segmentation methods, clustering based methods have been widely used for their simplicity, easy operation, robustness. Clustering based methods mainly rely on the some features which are extracted from original images, then take accord to different clustering thought to classify the feature space, at last, mapping to the original images to complete the image segmentation. Furthermore, different clustering methods can be chosen on request in the practical application.Gaussian Mixture Models(GMM) based Expectation Maximization(EM) Algorithm is one of the most used clustering methods in the image segmentation. As an unsupervised learning algorithm, it uses maximum likelihood method to estimate mixture parameters and can be adopted in many practical situations. The distribution of gray level and other features can be approximately regarded as GMM, so EM algorithm can be adopted in the image segmentation. But EM algorithm has its inherited weakness that it is sensitive to the initialization parameters and easily traps into the local optimal points; while the reconstruction of a signal in the scale-space theory can acquire the clear distribution of features so that it can be helpful to obtain the initialization parameters that EM needs to improve the segmentation quality. In this paper, we proposed an improved image segmentation method based on GMM based EM and scale-space theory to improve the robustness and accuracy of image segmentation so that we can achieve the aim of adaptive segmentation.Three aspects are included in this paper:(1)The theory and development of EM clustering algorithm and scale-space methods are introduced, EM, its invariant algorithms and the theory of reconstruction of a signal and fingerprints theory are illustrated in detail in this paper.(2)To alleviate the disadvantage of traditional EM algorithm, a novel improved EM method based on the theory of scale-space has been introduced when we consider only one feature in image segmentation. In this method, firstly, estimating the density distribution of the feature; secondly, obtain the adaptive parameters that EM needs based on scale-space theory; lastly, segment images based on the initialized EM. The experiment on true brain MRI suggested that our method can achieve the best results compared with the segmentation groundtruth.(3)A novel improved EM segmentation method based on the more than one image feature has been proposed. The true distribution of high-dimensioned data set and the influence of noise are both considered in the method, the initialization information for EM is obtained by screening the true clustering center. Experimental results in the segmentation dataset from California University showed that our method is superior to other EM invariant algorithms in the two clustering criterions called Xie–Beni and Q.