Research On Improved Image Segmentation Method Based On Spectral Clustering

In recent years, the image segmentation method has been paid more and more attention, because the image segmentation process is the key step between image processing and image analysis. Spectral clustering method is a popular image segmentation method in recent years. Because it is not restricted by the shape of the sample space, and can be clustered in any sample space. In spectral clustering algorithm, the structure of the similarity matrix is the key of the algorithm. The similarity matrix directly affects the segmentation results of the algorithm. However, there are some problems should be solved such as inaccurate segmentation results with the error parameters and low efficiency of segmentation because of mass data operations. Therefore, the research of similarity matrix is very necessary in spectral clustering algorithm.In this paper, the similarity matrix of spectral clustering algorithm is mainly studied in the image segmentation algorithm based on spectral clustering. The specific contents are as follows:(1) Research on the construction of similarity matrix in spectral clustering image segmentation.When the color images are segmented by traditional spectral clustering algorithm, the only one of color space and distance calculation formula is usually used to construct similarity matrix. The influence of the segmentation results which established on the different color space and distance calculation formula is neglected. It leads to many limitations of spectral clustering algorithm. To solve this problem, using Euclidean distance, cosine distance and chi square distance formula, the different similarity matrices are established on RGB and HSV color space. The best segmentation construction method of the similarity matrix is obtained by analyzing and comparing the effect of different construction methods. The effectiveness of spectral clustering algorithm for segmenting color images is improved. By calculating the accuracy of image segmentation results and the performance evaluation index of precision and recall, the reliability and accuracy of the experiment are verified.(2) Research on the improvement of similarity matrix in spectral clustering image segmentation.The Gauss kernel function based on Euclidean distance is generally used to measure the similarity between the sample data points in traditional spectral clustering image segmentation algorithm. Because only the local consistency is taken into account,and the scale factor of Gauss kernel function needs to be set up manually according to the experience, so that its accuracy can not be ensured. Aiming at these defects, a new similarity matrix construction method based on spectral clustering of weight modified cosine distance is presented, which avoids setting scale factor in traditional spectral clustering algorithm. The limitation of traditional cosine distance in which only the similarity of vector dimension direction is taken into account is eliminated. In the spectral mapping process, using Nystr?m approximation method, the characteristic value and characteristic vector of similarity matrix are approximated, so that the computational complexity which solves similarity matrix is reduced when the similarity matrix is solved. The segmentations results are compared with traditional spectral clustering algorithm and the algorithm of traditional cosine distance to construct similarity matrix by the images of Berkeley image database segmentation experiments. The results show that the accuracy and the segmentation effect of the algorithm in this paper is the best in three spectral clustering image segmentation algorithms.