Research And Application On Adaptive Spectral Clustering Algorithm

Spectral clustering is one of the most popular modern clustering algorithms. However, the performances of spectral clustering are severely dependent on the value of the scaling parameter. To the best of our knowledge, choosing an appropriate scaling parameter for spectral algorithm is still a difficult problem. In this thesis, we propose an adaptive spectral clustering (ASC) algorithm which can effectively deal with the parameter selection problem. Furthermore, we extend ASC for semi-supervised cases and propose a semi-supervised adaptive spectral clustering (Semi-ASC) and a semi-supervised spectral ensemble clustering algorithm. At last, we apply our algorithms to the field of image segmentation. The main contributions of the thesis are summarized as follows:Firstly, we apply the kernel selection trick to spectral clustering algorithm to solve the parameter selection problem of spectral clustering algorithm and propose the ASC algorithm. We compare our algorithms to existing parameter selection methods for spectral clustering algorithms on both synthetic and UCI datasets. Experimental results validate the effectiveness of the proposed algorithms.Secondly, we extend ASC in the semi-supervised learning area and propose semi-supervised adaptive clustering algorithm (Semi-ASC). Furthermore, ensemble trick has been applied in the S-ASC algorithm and we propose semi-supervised spectral ensemble clustering algorithm. Experimental results on both synthetic and UCI datasets validate the effectiveness of the proposed algorithms.Thirdly, we apply our algorithms in the image segmentation area. Some important image pre-processing steps are taken to filter and smoothen the image. We also modify the distant measurement of the spectral methods to meet the requirements of image segmentation. Experimental results on Berkeley Segmentation Dataset validate the effectiveness of the proposed algorithms.