Cluster-based Image Segmentation And Classifier Design

Traditional pattern recognition involves two important tasks: unsupervised clustering and supervised classification. The unsupervised clustering methods can group samples into meaningful clusters, and thus make an assessment of data structure and further facilitate better understanding for the data. The supervised classification methods can utilize the class information to determine their decision functions which can provide a class label to a newly encountered, yet unlabeled sample.The dissertation consists of two parts. Firstly, when class labels of given samples are unavailable, the effective clustering methods are developed to segment the gray-value images in an unsupervised manner. Secondly, when class labels of given samples are available, the methods realizing both clustering and classification learning are presented to not only reveal the underlying structure of the data but also construct the effective classification mechanism. The main contributions of this dissertation are summarized as follows:(1) A fast and robust fuzzy c-means framework is constructed for image segmentation. This framework has two major characteristics: (a) a new factor Sij as a local (both spatial and gray) similarity measure is designed to guarantee both noise-immunity and detail-preserving for image; (b) the fast clustering is adopted to make the segmenting time only dependent on the number of the gray levels Q rather than the size N (>>Q) of the image, and thus reducing the computational complexity significantly. We utilize the typical clustering methods Fuzzy C-Means and Gaussian Mixture Model to demonstrate the feasibility of our framework.(2) A robust fuzzy relational classifier is developed to handle the datasets containing outliers or non-spherical structures. It can not only uncover the data structure,but also provide the class label for unseen samples. Furthermore, it can utilize the relationship between data structure and class to interpret the classification result.(3) A simple enhanced fuzzy relational classifier is presented. By employing the training samples differentiatedly to build a more robust and effective R, this classifier achieves three advantages: robustness for classification, effectiveness for classification and low computational load for the construction of R.(4) A framework for simultaneous clustering learning and classification learning (SCC) is proposed. The contributions of this framework are: (a) designing an effective classification algorithm more prone to be transparent; (b) robustly clustering the data with guidance of supervision information to form relatively pure clusters; (c) adaptively revealing a statistical relationship between clusters and classes. Based on SCC, a multi-objective framework for simultaneous clustering learning and classification learning (MSCC) is developed. Compared to SCC, MSCC has two advantages: (1) the parameterβin SCC can be avoided by utilizing the multi-objective functions to describe the clustering problem and classification problem, respectively; (b) the effectiveness of SCC can be improved by extending the solution space of SCC.(5) A general framework for simultaneous clustering learning, classification learning and metric learning (SCCM) is presented. Here the metric learning is equivalent to the feature-weight learning. Therefore, this framework can achieve four goals: (a) learning feature weight to reflect the importance of the feature; (b) realizing the effective clustering in the linear-weighted feature space; (c) designing the effective classification mechanism based on the structures in the new feature space; (d) uncovering the statistical relationship between the structures in new feature space and classes in the output space.