Semi-supervised Mean Shift Framework And Its Application Of Image Segmentation

Semi-supervised learning has become one of the hottest topics of machine learning recently. As a branch of the semi-supervised learning, semi-supervised clustering also has drawn more and more attention. Semi-supervised clustering can use a small amount of supervised information to guide the clustering algorithm towards a better grouping of the data or to do metric learning. So compared with traditional clustering algorithms, Semi-supervised clustering can usually get a better result. Previous semi-supervised clustering methods were usually developed by modifying their unsupervised counterparts, introducing punishment items into the original clustering objective function to solve the problem of violation of constraints. Although mean shift is one of the most popular algorithm in the field of computer vision, it is difficult for the mean shift to be converted into a semi-supervised form using the above method directly. It may be the main reason that there is less research on semi-supervised mean shift.This dissertation focuses on modifying mean shift to its semi-supervised form, puts forward a new framework, and then applies the semi-supervised mean-shift model to image segmentation. The primary work of this paper can be summarized as follows:1. Base on a comprehensive summary of the mean shift algorithm, resorting to the idea of pairwise constraint propagation (PCP), we make the first try to obtain a pairwise constraint based semi-supervised mean shift (SSMS) by defining an equivalent objective function. And the constraints can also help choose bandwidth in the mean shift. The experiment on artifical and UCI data sets verify the effectiveness of this algorithm.2. In image segmentation, using a small amount of hand-marked pairwise constraints can not only get the needed bandwidth, but also obtain a better image segmentation results.3. Running mean shift algorithm directly uses the histogram of gray image. And this significantly reduces the computational complexity. Further, we develop a fast model for gray image segmentation through transforming iterative solution into non-iterative one.