Reconhecimento de caracteres manuscritos baseado em regiões perceptivas

AUTOR(ES)
DATA DE PUBLICAÇÃO

2005

RESUMO

This work investigates the perceptual zoning mechanism for handwritten character recognition. It is proposed a non-symmetrical zoning mechanism as the baseline on the analysis of the confusion matrix for each individual classifier (Class-Modular). Zoning is a method for local information analysis on partitions of a given pattern. The feature extraction is based on Concavities/Convexities deficiencies, which are obtained by labeling the background pixels of the input images. Therefore, circumscribes the letter by a rectangle and partition it into Z parts, such as: Z = 4, 5H(horizontal), 5V (Vertical) and, 7 parts. The base of data used for the experiments is IRONOFF, with handwritten characters of the alphabet. For the recognition problem a Neural Network team is proposed, where the K-classification problem is decomposed into K 2-classification sub problems, each for one of the K classes. A methodology for multiple classifiers system (MCS) is applied to the recognition problem, could be used for the fusion (combination) of classifiers. The methodology defines an alternative approach instead of using the recognition rate criterion, which can be used to evaluate a priori classifiers combination in MCS. The obtained recognition rate for the evaluated zonings are the following: 4 = 82,89%, 5H = 81,75%, 5V = 80,94% and 7 = 84,73%. The combinations accomplished among the individual classifiers present an improvement in the rate recognition, being the best result of 85.9% for the network 5H-5V-7. The global result considering a composed architecture for 2 classification levels (meta-class and class) reaches an average recognition rate of 84,15%, with rejection of 11,95% and error of 3,90%.

ASSUNTO(S)

redes neurais (computação) computação gráfica informática ciencia da computacao processamentos de imagens

ACESSO AO ARTIGO

Documentos Relacionados