Sistema de reconhecimento de palavras manuscritas dependente do usuário.

Abstract

This work presents a writer-dependent system for isolated handwritten cursive word recognition. This system is characterized by the utilization of a pre-processing state, which corrects imperfections and normalizes variations in the word image, an explicit segmentation stage, which splits the word into characters or character segments, a feature extraction stage, which represents the image by three feature vectors (perceptive, global and directional features), and a vector quantization module, which performs the mapping of a feature vector into an observation vector (or symbols vector). The symbols correspond to indices (the code vectors) generated by the representation (vector quantization) of the feature sequences with the use of dictionaries. Finally, there is the classification stage, performed by Hidden Markov Models, where characters are individually recognized and combined to form a valid word. Experimental tests were conducted with a database specifically built for this problem, containing samples of manuscripts from 4 different writers. The writer-dependent system for isolated handwritten cursive word recognition was recognition rate between 83.31% and 92.96% depending writer analyzed. The results show that the system offers optimum performance when used word recognize by the characters models.