Visual information extraction: Region-of-interest detection, digital Zernike moments and multi-point descriptors

Automated visual information extraction is an important and challenging task for computer vision. This dissertation addresses visual information extraction from two specific perspectives: (1) with some prior information, how to detect the existence and find the areas that contain the targets of interest in an image; (2) without any prior information, how to find correspondence in image sequences?; The dissertation first introduces a region-of-interest (ROI) detection algorithm. This algorithm aims at quickly detecting and isolating regions in an image that contain portions of a target. The algorithm makes use of prior information of target template and image resolution to derive a hierarchical, multi-scale architecture. Morphological processing is developed to improve the chance that each ROI contains only one target. To alleviate the assumption of having the image resolution known a priori, an image resolution estimation algorithm is developed based on camera conditions, such as field-of-view angles and orientation of the camera.; The second image information extraction issue addressed in this dissertation deals with rotation invariant moments for feature based target recognition. Even though the Zernike moments have been successfully applied as a transformation for continuous variables, the artifacts brought about by digitization for image processing application have not been systematically addressed. In this dissertation, the lost orthogonality due to digitization of the Zernike moments is uniquely formulated as an optimization problem. A numerical procedure is then developed to solve for the new Zernike moments.; The third contribution of the dissertation is concerned with video images. It focuses on developing new image descriptors in an image sequence. Unlike traditional approaches that utilize a single control point for matching pairs of points, the new descriptors are defined on multiple points. The multi-point descriptors are shown to be more reliable under image transforms, more accurate in finding corresponding control points, and capable of matching multiple pairs of points simultaneously. This algorithm is then applied in an image mosaicking system. It is evident that the mosaicking system is robust in response to target motion, luminance change, and other noise, and capable of adjusting computational complexity according to the image content.