Image segmentation and classification are the two main issues in image processing and computer vision. Segmentation deals with the problem of separating an image into a number of distinct regions that correspond to objects/surfaces in the original scene. Classification is the process by which an image (or image region) is identified as being one among a group of patterns/models known to the system. This dissertation combines color and texture that are available in an image, and constructs a unified framework for answering both of the above questions. Color in texture analysis has largely been ignored, while the majority of proposed methods involve only the intensity information (luminance) of an image. A set of real-time, computationally efficient and easily implementable algorithms is designed and implemented for both segmentation and classification. The segmentation system processes luminance and chrominance separately and combines the results. Luminance processing follows a three-step procedure: (a) filtering, (b) smoothing, and (c) boundary detection. Chrominance processing involves two main steps: (a) histogram multi-thresholding, and (b) Region Of Interest (ROI) expansion. By combining the results from luminance and chominance, a methodology for detecting, locating, and measuring image changes in the ROI is developed. The classification system proceeds in four stages: (a) computation of autocorrelation and cross-correlation of both luminance and chrominance, (b) extraction of directional histogram features, (c) statistical selection of robust features, and (d) neural network classification based on the selected features. The proposed system is based on the xyY color space and proposes a two channel vision system based on a chromaticity mapping compression scheme which is efficient while producing negligible loss of chromatic (color) information. The system also implemented in the HIS space. Both spaces are proven to provide computationally superior systems compared with the RGB color space. Furthermore, a multiple-threshold segmentation method, based on the Total Color Difference (TCD) measure, is also developed. Experimental results justify and support the use of color in addition to luminance. The end result of the proposed work is the design of a complete Visual Monitoring System (VMS) for automated surveillance, which is capable of detecting and identifying potential changes in the surveyed environment over a period of time, with applications in wetlands monitoring, Autonomous Underwater Vehicles (AUV), and Geographical Information Systems (GIS).
展开▼
