Image character through signal and pattern formation

Optical Character Recognition (OCR) is one of key research areas of Artificial Intelligence (AI), and image text recognition is one of challenging fields of OCR. Presented work offers a character recognition system for cursive script (e.g., Arabic, Urdu, etc.) segmented characters from their images. Presented methodology consists of phases namely (1) Image Acquisition, (2) Preprocessing, (3) Chain Code Formation, (4) Signal Generation, and (5) Pattern Extraction. Image Acquisition takes an image of segmented character and converts it into binary image. Preprocessing finds out unconnected regions from the binary image and separates into one main region and one or more secondary regions. The secondary regions make vector of features whereas main region is further converted to one pixel wide thinned image. Chain Code Formation finds out chain code vector using 8-directions. Finally Pattern Extraction uses a defined algorithm to form qualitative signal pattern describing increase (+1), decrease (-1), and constant (0) pattern of writing layout. Additionally recognition of the character is carried out through Feed-Forward Neural Network (FFNN). The methodology and presented algorithm are evaluated using 1292 images of segmented characters of 271 different ligature classes of printed script. The methodology is tested on MATLAB and overall recognition rate obtained with Mean Squared Error (MSE) of 0.0014 with 60 hidden neurons of FFNN.