Novel structure-recognition-based OCR system and its parallel VLSI architecture
Tie-Jun Shan · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1993
We propose a structure pattern recognition based OCR system for printed text recognition. The segmentation algorithm is a raster-scan based one that reduces memory access time a great deal. The feature extract algorithm is a chain-code encoding based graph traversal algorithm. It has an advantage of single memory access for each pixel while most graph traversal algorithms require multiple scans of the entire image. While traversing a thinned character, structural feature is automatically extracted. The classification is a process of tree search in the structure feature space that is created during the training process. The segmentation, thinning, and feature extraction algorithms are all raster-scan based algorithms that can be implemented by a parallel systolic array architecture.