Results:
- The probabilistic model of combining character matching scores was constructed.
- The existing handwritten word recognition algorithm was modified to use this model resulting in improved performance.
In this project I considered handwritten word recognition based on segmentation of word images into character images, recognizing character images and combining results of the recognition. The goal of this project was to derive a best way of combining character matching scores into word recognition scores.
Tulyakov, S. and V. Govindaraju (2001). Probabilistic Model for Segmentation Based Word Recognition with Lexicon. 6th International Conference on Document Analysis and Recognition (ICDAR 2001), Seattle, USA, IEEE Computer Society.
In this project I was trying to locate a destination address block from binarized pre-segmented images of US mailpieces of low quality. The task was complicated since images did not necessarily contain needed address block, or sometimes contained a multitude of extraneous information and noise. The idea of the algorithm was to extract contours and to map contour information into special feature-position space. After performing clustering in this space, a candidate clusters corresponding to address blocks were selected and their properties verified. The algorithm performed robustly, and was able to increase the overall address recognition system performance by 10%.
Govindaraju, V. and S. Tulyakov (2003). Postal address block location by contour clustering. 7th International Conference on Document Analysis and Recognition (ICDAR 2003), Edinburgh, Scotland, UK, IEEE Computer Society.