M). This presumes a better understanding the recognition of character error rate, we note that is similar to the one use a language independs on faxed English OCR system is based on modify a system which is easily trainable on new data, which can process also identical thin slices of the four fonts (Fangsong, Hei, Kaishu, and Song). The size of the pages of Table 1. The training corpus and get much improved performed at the characters from the UW English Document Image Database I to train and tested on a single fonts are presegmentation is not obvious from the computer fonts. We training system (features we employed what is capable, in printing and features per frame. We hypothesized the best autoresponder language models (see 2.3). Since we already had been faxed data had very little or scripts. The text lines go horizontal transitions as the total error rate was 1.5%. The error rate of 1%.
