Toward a robust and accurate vision system which recognizes character images.

摘要

Substantial research effort have been devoted to the development of automatic character recognition, which has led to the blooming of office automation industry.; Recent developments have concentrated on solving problems due to seriously degraded images. To deal with these problems, algorithms of character recognition technique at various functional levels have to be redesigned or enhanced. In this thesis, we target at developing new character recognition techniques for preprocessing, feature representation, recognition and pose estimation, so as to handle degraded character images.; For image preprocessing, a novel algorithm for character region extraction based on grayscale mathematical morphology is proposed. It detects features that may be occurred rarely in other regions, except the character regions. Under complex scene and noisy environment, character regions in 36 out of 40 natural scene pictures are correctly exacted. Based on the rough set theory, these extracted character images can be further manipulated to generate skeletons of good quality in one pass of scan, which are useful for syntactic and structural recognition.; In order to represent characters in consideration of perspective distortion due to different viewpoints, an approach using Fourier-coefficients based descriptors is presented. The approach is proved to be viewpoint invariant and more efficient in implementation as compared with other techniques. In addition, a wavelet-transformed based semi-local descriptors are proposed to describe the boundary of degraded character images. These descriptors are better than other global descriptors such as Fourier descriptors and moments as they can include both global and local information on the shape.; For recognition, a new method of similarity measure for Fuzzy-attributed graphs (FAGs) is proposed which provides a more robust tool to compare degraded character images from their reference patterns. Experimental results show that an accuracy rate around 93 percent can be achieved by our approach for handwritten numerals.; Finally, a direct solution of pose information including viewing distance, rotation angle, slant and tilt angles for known characters is proposed. It provides a more accurate recognition or simpler segmentation by feedbacking these information to adjust the position and orientation of the camera. It has been demonstrated that this method is accurate, efficient and robust to high frequency noises and even small occlusion. (Abstract shortened by UMI.)