Image Segmentation Via Spectral Clustering With Manifold Structure And Kernel Propagation

This thesis mainly deals with image segmentation which has been received much attention in the machine learning community. The main contributions include image segmentation based on spectral clustering with manifold structure, image segmentation via kernel propagation and interactive foreground extraction using kernel propagation.In data clustering, spectral clustering has been effectively and widely employed in many research fields. However, when it is applied to image segmentation, it causes high computational complexity and instable results and needs large storage space. Image segmentation based on spectral clustering with manifold structure is proposed. The proposed algorithm exhibits a high-performance, not only in reducing computational complexity and storage space, but also in avoiding instable results.Kernel propagation (KP) is an effective semi-supervised kernel matrix learning (SS-KML) method which improves the comprehensive performance. By applying KP to image segmentation, we design a new image segmentation method in natural images. Global k-means clustering and self-tuning spectral clustering are employed to select globally-optimized seeds which are used to generate the seed-kernel matrix. Finally, we propagate the seed-kernel matrix into the full-kernel matrix of the entire image, and thus image segmentation results are obtained. As a result, our method outperforms state-of-the-art approaches for image segmentation.We propose an effective interactive foreground extraction method by applying KP to interactive image segmentation. Mean-shift segmentation is used to generate superpixels, and then the users simply draw strokes (which called markers) to indicate the seeds of the foreground and background. We generate pairwise constraints from the seeds to make the seed-kernel matrix for KP. Next, we propagate the seed-kernel matrix into the full-kernel matrix of the entire image using KP. Experimental results show that the proposed method preserves statistical characteristics of the seeds during the propagation of seed-kernel matrix by KP and achieves the state-of-the art performance.