BPCC-based approach for arabic letters recognition
Last modified: April 25, 2006
This paper introduces a method of Binary Pairs and Chain Coding (BPCC) using important points for Arabic letters recognition. It is a good method paid to decrease the size of data at least 38%, reduce the number of calculations and save the time. The important points are detected. These points are coded through two counterclockwise passes; diagonal then perpendicular depending on a novel algorithm. Eight sets of binary pairs are supposed and they are used to filter the resulted code of character. Filtered code is dealt with a certain method to form an array of 85 elements. Then the array is scaled to consider as an input of NN.
Neural Network consists of one input layer of 85 neurons, one hidden layer of 500 neurons and output layer of 16 neurons. Training parameters of NN are set to the default values, momentum coefficient is 0.25, learning rate is 0.05 and the sum square error is 0.00000585. The NN uses bipolar function as an activation function. Once the NN weights and biases are generated randomly, the range of input vector values is ranked between -1 to +1.
Two types of Arabic letter fonts of multi size are used; they are Arial, and Simplified Arabic. The results prove the recognition rate is 99.3% for Arabic Letters recognition of two fonts of multi size. The method is easy implementations, reliable to gain fast speed, high competence, hopeful results and encouraged.