We present partial results of a project to build a digital collection of a very important set of documents for the History of Mexico: a collection of telegrams written by Porfirio Diaz, a president of the country who ruled at the beginning of 20th century. This article focuses in the recognition of isolated words taken away from telegrams written by Diaz. Images of the telegrams were obtained from microfilms of the originals, hence there is a lot of noise and poor resolution on them. This, added to the fact that old handwritten manuscripts are hard to read, make this task especially difficult for a automatic recognizer. We designed a recognizer based on a back-propagation neural network, that works recognizing each letter in each word. The recognizer receives as input a cleaned word and finds the class of each character in the word. The pre-processing on the words and the segmentation algorithm to get each letter is also described.
展开▼
