Unsupervised Segmentation Of Images Based On Spectral Clustering

Image segmentation denotes cutting image into several similar, uniform and connective parts in visual. It is one of the important steps in the area of image processing, analysis and recognition, also the key technique in computer vision. There are many applications of segmentation in industry, agricultural, military, medical image, Internet and robot vision. According to with human interaction or not in process of segmentation, there are three kinds of image segmentation methods:unsupervised, semi-supervised and fully supervised methods. Among many image segmentation algorithms, unsupervised image segmentation is one of the difficult and important methods. Therefore in this paper, it mainly focuses on the unsupervised image segmentation, proposes a spectral clustering based unsupervised segmentation method, and carry experiment on Berkeley Segmentation Database.At the beginning, this paper specifically introduces the basic ideas in image segmentation, gives out the backgrounds of segmentation and the mathematic definition of image segmentation. Then it focuses on the unsupervised segmentation algorithms. Among these current classic unsupervised segmentation methods, it mainly talks about the ideas of Normalized cut, Mean shift, Felzenszwalb-Huttenlocher, Compression-based Texture Merging, Total Variation Segmentation, Learning Full Pairwise Affinities, Gpb-owt-ucm.Secondly, this paper details the framework of the proposed algorithm. It first introduces the concept of superpixels, and then integrates geodesic edge, position and color features to form the similarity matrix W, so that it more accurately describes the similarity between data. After that, this paper uses spectral clustering on the similarity matrix constructed in the last step to complete the image segmentation. There are two innovations in this paper, the first one is using superpixel instead of pixel to reduce the amount of data, the second one is the use of geodesic boundary cue instead of a straight line boundary cue.In the end, this paper introduces current image segmentation evaluation criteria and makes some analysis on these criteria. Then it carries full experiment in the Berkeley Segmentation Database, reaches relatively good segmentation results, while the results were also evaluated with criteria and compared against current classic methods. Experiments show that this algorithm significantly performs better than the current classical algorithms except Gpb-owt-ucm.