Study On Image Segmentation Based On Clustering Analysis

Image segmentation is the method that extracts interesting information in image, and has become the key step from image processing to image analyzing. After developing and changing in a period of decades, various image segmentation methods based on various theories have been proposed. In which, the image segmentation method based on clustering analysis was born in the development of clustering analysis. Clustering analysis includes two classes:the one is hard-partition clustering; the other is soft-partition clustering. For the image segmentation problem, the core idea of them is to partition pixels into clusters based on the similarity between pixels. The relationship between them is complementary. The image segmentation method based on hard-partition clustering analysis shows high computational efficiency on solving image segmentation problem having simple content. The image segmentation method based on hard-partition clustering analysis can solve the image segmentation problem having complex content, which the former solves hardly. It remedies the defect of the former.However, for clustering algorithm, there are still some problems which influence the computational efficiency and the segmentation effect of image segmentation. In this paper, the computational efficiency of the hard-partition clustering algorithm and the segmentation effect of the soft-partition clustering algorithm are analyzed and researched respectively.For the hard-partition clustering algorithm, to avoid falling into local minima during searching the local optimal partitions, a lot of related researchers have proposed various methods. J-means algorithm is a local search heuristic that have more power to escape from local minima, which has been verified by experiments. Compared to others local search heuristic, it can produce a less clustering error. But, it costs more computational complexity. Especially, for image segmentation in real time, it can not satisfy the demand of computational efficiency. Thus, in this paper, a fast J-means algorithm is proposed, which uses the minimum intra-clusters or the maximum inter-cluster variance of subpartition to constraint the range of searching a local optimal solution. A candidate neighboring solution is produced in all neighboring solutions before searching local optimal solution, then, the local optimal solution is found out from this candidate neighboring solution. The proposed algorithm is developed from formula and is proved theoretically, which can decrease the computational complexity and increase the computational efficiency. In experiments, standard data sets and real airborne remote images are employed to test the proposed algorithm. The experimental results demonstrate that the fast J-means algorithm saves a half of the time used by J-means algorithm at least based on a similar or an identical result. Especially, in the experiment of the real airborne remote images, the fast J-means algorithm satisfies the demand of image segmentation in real time.For soft-partition clustering algorithm, the finite mixture model applies Bayesian rule to partition pixels into various clusters, which supplies an efficient mathematical method that simulates the complex probability density function by simple probability density function. It attracts much attention in various fields. In the recent decade, some related literature at home and abroad pointed out that, because the finite mixture model doe not incorporate the spatial relationship between neighboring pixels into the prior, the image segmentation results are sensitive to noise. Thus, spatially various finite mixture model was proposed, which firstly introduces the spatial relationship between neighboring pixels. In the background of this model, many related researches are developed to enhance the spatial information in the prior, which reinforce the power of noise suppression. In this paper, a new spatially variant finite mixture model is proposed, the spatially variant finite mixture model based on morphological dilation. In the model, a posterior probability function is defined for each pixel, and it is processed by the redesigned morphological dilation function for neighboring pixels having a same posterior probability function. Such method incorporates the spatial relationship between neighboring pixels into the prior of each pixel, which enhances the capability of image segmentation against noise. For verifying the performance of the proposed model, example image, synthetic image, medical MR image and medical CT image become experimental objects. The experimental results illustrate that the proposed model generates more clustering accuracy and less sensitiveness to noise compared with other spatially variant finite mixture models. Moreover, the fixed spatial neighborhood of spatially variant finite mixture models makes the spatial information incorporated into the prior adapt the local space weakly. In this paper, in the background of the spatially variant finite mixture model based on morphological dilation, a spatially variant finite mixture model based on variant spatial neighborhood is developed. In the proposed model, a spatial neighborhood pattern is defined, which contains edge pattern and non-edge pattern. For each pixel, the pattern is distinguished by a LVQ artificial network. For the pixels belong to edge pattern, their spatial neighborhood remain unchanged, for the pixels belong to non-edge pattern, their spatial neighborhood are enlarged. Such processing make the prior adapt local space in some sense. Experiments adopt example image, synthetic image and real heat infrared image to test the proposed model. Compared to the model with the fixed spatial neighborhood, the proposed model has stronger inhibition ability on a higher level noise.