Research Of Image Segmentation Algorithm Based On Density Peaks

In image processing, image segmentation is an important analytical technique aimed to divide image into regions with different characteristics and extract meaningful image region, it is the basis for the subsequent image processing. In the field of computer vision, image segmentation has been a hot topic of research, there has been endless variety of ways to solve the image segmentation in a variety of complex issues, but enhancing the result of image segmentation is very slow, how to segment image accurately and efficiently becomes the challenging work in this area.This thesis gave a brief introduction about supervised and unsupervised image segmentation algorithm for image segmentation algorithm research status, and made a detailed description of the classic unsupervised image segmentation algorithm. The existing unsupervised image segmentation algorithm has the following two common problems:(1) There are limitations in the segmentation of complex shape of the object. Clustering algorithm based on partition or hierarchical can only find the "class round" clusters. Therefore, when they applied to image segmentation, the segmentation effect is poor.(2) Unsupervised segmentation algorithm is sensitive to texture information of image. On the detection of the edge, unsupervised segmentation algorithm usually detected the texture of the object as the edge.For the problem of density clustering, this thesis fused kernel density estimation and the peak density clustering, and proposed a new unsupervised image segmentation algorithm based on the density peaks(DP-UIS). With respect to the existing work, the main work is as follows:(1)This thesis proposed a new unsupervised image segmentation algorithm based on the density peaks(DP-UIS), it can deal with the nature image segmentation with complex texture and shape effectively, and ensured that the natural images were segmented in real time.(2) Based on the color, coordinate and texture information, this thesis defined the metric function of the density difference, and introduced the density difference threshold, thus, DP-UIS can determine the number of divided regions automatically.(3) Since the density peaks clustering algorithm was proposed, it just dealt with the clustering of data points. This thesis extended the improved peak density clustering to the field of image segmentation to obtain a better image segmentation results.This thesis verified the validity of the DP-UIS algorithm by Berkeley image database. Compared with the classical Mean-shift algorithm, K-means algorithm, FCM algorithm and N-cut algorithm, it obtained a better image segmentation results. In addition, this thesis analyzed the parameters used in the experiment and described how the changes of the parameters to effect the image segmentation results in detail.