Video-based handwritten Chinese character recognition

In this thesis, we develop a novel Video-based handwritten character recognition (VCR) system focusing on the recognition of Chinese handwritten characters. The main problem of VCR system is how to effectively extract stroke dynamic information from video data for character recognition. We design a stroke extraction algorithm that utilizes the combination of stroke static information analysis and dynamic information analysis. Static information analysis relies on the static/spatial information of character image to extract the character stroke. Based on the result of static information analysis, dynamic information analysis is then performed to extract all the dynamic/temporal information of character strokes. The recovered character information is analogous to that of on-line systems, so a conventional on-line method can be utilized for the character recognition.; To extract the character static information, we propose a novel off-line Chinese stroke extraction method. We investigate all the skeleton points that exist in a skeleton character image, and improve the feature point detection method based on Rutoviz's crossing number definition. Using a new bi-directional graph, we successfully connect Chinese character stroke segments. The algorithm can accurately extract the strokes from the thinned Chinese character images. Extensive experimental results on over eighteen thousand character strokes show our method can achieve over 99% accuracy.; For the dynamic information extraction from video character sequences we propose two schemes, pixel-based scheme and stroke-based scheme. The pixel-based scheme extracts the pixel-based dynamic information with enhanced noise-resistance capability. Since a pixel-based analysis method has to consider too many complicated cases that might degrade the system performance, stroke-based scheme is proposed to effectively extract the character dynamic information. We find a dynamic clue in image sequence that can help to recover the dynamic information for each static stroke. Based on this a mathematical model is built to analyze the stroke's drawing slot. A powerful error-correction capability is also provided that can automatically verify and correct most of errors introduced by the static analysis. The experimental results on over 3000 video character sequences show that our system can extract the Chinese character stroke dynamic information robustly. The extracted stroke information has been used for Chinese character recognition via the attributed relational graph (ARG) matching method. About 99% of the strokes were extracted correctly, and 97% recognition accuracy is achieved, the performance is comparable to that of the corresponding on-line system. (Abstract shortened by UMI.)...