Cursive Script Postal Address Recognition Abstract Cursive Script Postal Address Recognition

Cursive Script Postal Address Recognition By Prasun Sinha Large variations in writing styles and di culty in segmenting cursive words are the main reasons for cursive script postal address recognition being a challenging task A scheme for locating and recognizing words based on over segmentation followed by dynamic programming is proposed This technique is being used for zip code extraction as well as city state recognition in our system Results have been reported based on cursive script images from the United States Postal Service USPS database An overview of the cursive script postal address recognition system under development at the IBM Almaden Research Center is presented Optical Character Recognition OCR systems are usually trained to minimize character level errors which does not necessarily guarantee the best results for word level recognition A scheme for combining character level optimized classi ers using a word level optimization function to achieve higher word level recognition rates is proposed The optimum values for the parameters of the combination function are obtained using the gradient descent method Improvements in word recognition rates on the USPS database by using word level optimized ensembles of classi ers are reported