Image Segmentation Based On Multi-features And Subspace Clustering

Image segmentation is the process of dividing the image domain into several non-overlapping regions, with each region having distinctive significance. It is a key step in analyzing and understanding an image, and has wide applications in computer vision and object recognition.Based on the fact that high-dimensional data usually lies in a union of several low-dimensional subspaces, subspace clustering methods segment high-dimensional data into the subspaces they belong to by combining subspace representation of high-dimensional data and spectral clustering. Image segmentation can be regarded as a clustering problem of image feature data, therefore the subspace clustering provides a natural way for feature based image segmentation. In recent years, subspace clustering methods have been widely applied in computer vision and machine learning, however its application in image segmentation is still at a preliminary stage. In this thesis, we consider mathematical models and algorithms for image segmentation based on image feature and sparse subspace clustering. In natural images, the objects may have various features such as color feature, texture feature and geometric feature, therefore a single type of feature is unable to accurately describe the diversity of natural images. We propose to jointly use multiple types of features.In this thesis, we improve the sparse subspace representation model by introducing two kinds of weight in the 1-norm of representation coefficients. One weight is related to the data similarity and the other is related to data correlation. The weighted 1-norm penalty of the subspace representation coefficients tend to force similar data or linearly related data be involved while dissimilar or unrelated data not involved in the linear representation of a datum. The resulted subspace representation can overcome the drawbacks of 1-norm penalty that the presentation coefficients are usually over-sparse and not robust for highly correlated data. We test the performance of our method on the MSR500 dataset which is commonly used in image segmentation community. We evaluate the segmentation results by visual and objective index. Experimental results and objective assessment indexes show that the proposed methods can obtain stable, accurate segmentation results, which are more consistent with visual perception than some related methods.