Real Coded Genetic Algorithm Based Neural Network Model for Odia Numerals Recognition

Character recognition has great importance in present scenario. It has many application areas in the field of business, postal system, banking, library, form processing, document processing etc. A highly efficient character recognition system is required for such type of applications. Various authors have used different classifiers, mainly based on neural networks for this purpose. As BackPropagation (BP) algorithm is a derivative based algorithm, the chances of the results to falling to local minima is there. To alleviate problem in this paper we have proposed a hybrid system for recognition of Odia numerals by using multi layer neural Network (MLNN) and real coded genetics algorithm (RCGA). As RCGA is a derivative free algorithm it will overcome the problem of trapping the results into local minima. And as we are using the real coded GA (Genetics Algorithm) it will be advantageous over the binary coded GA, as we do not have to do the conversion from binary to real each time which saves the training time. Real coded chromosomes are used by GA to determine the weights of Neural Network (NN). Before recognition, preprocessing, feature extraction and feature reduction steps are carried out. For feature extraction Gradient based approach is used. The gradient of the images are calculated by applying Robert’s filter and the feature vector is generated. After the generation of feature vector PCA (Principal Component Analysis) is applied to reduce the size of features. The proposed system is applied on the standard dataset taken from ISI Calcutta containing 1200 samples of Odia handwritten numerals. From experimental result it is observed that the proposed system has achieved 98.33% accuracy on test dataset.