Abstract

Theories related to hierarchical data processing of soft computing are hot topics in the past several years, such as granular computing, multi-resolution analysis, multi-scale analysis, etc. By representing the universe of the problem in different granulation (resolution, scale), it provides a possibility to solve problems more easy, make problems solvable which can not be solved in the original universe, and reduce the computational complexity. In this dissertation, we study the application of hierarchical data processing to pattern classification and visual navigation. Results of the research are helpful to large scale data classification and analysis the content of image.

The first part of this dissertation discusses the relationship between multi-granule representation and processing with pattern classification. Research results can be summarized as follows. (1) By analyzing four kinds of risk minimization principles, the relationship between the size of the training data set and the trained classifier, and the possibility of combining the pattern classification with hierarchical data processing, it shows that the classification problem may be solved efficiently by representing the data set in coarse granulation. (2) An area based risk minimization principle is proposed. By representing the samples set by hyper-sphere (or hyper-cube), the boundary location can be controlled by the center and radius of the area. Experimental results show that the method proposed in this paper can reduce both the number of the training samples set and the support vectors. (3) We propose a multi-resolution classification strategy based on the partition of feature space. The training algorithm locates the boundary between two classes from coarse to fine resolutions by dividing the hyper-cuboids that lie on the boundary step by step. The testing algorithm firstly labels the testing data set by the classifier trained at initial resolution. Then only those lying on the boundary will be labeled at the finer resolution. Theoretical analysis and experimental results have substantiated the effectiveness of the proposed method.

  The second part of this dissertation discusses a road detection algorithm based on quotient space synthesization. This algorithm is composed of two modules: boundaries are first estimated based on the intensity image and road areas are subsequently detected based on the full color image. In the first module, an edge image of the scene is analyzed to obtain the candidates for the left and right road borders and to delimit the area that will subsequently be used to compute the mean and variance of the Gaussian distribution, assumed to be obeyed by the color components of road surfaces. The second module effectively extracts the road area and reinforces boundaries that the most appropriate fit the road-extraction result. The combination of these modules by quotient space synthesization can overcome basic problems due to inaccuracies in edge detection based on the intensity image alone and due to the computational complexity of segmentation algorithms based on color images. Experimental results on real road scenes have substantiated the effectiveness of the proposed method.

Keywords: granular computing; multiresolution analysis; pattern classification; support vector machines; road detection