Machine Printed Character Recognition System Using Feature Point Extraction And BP Neural Network Classifier

Optical Character Recognition has become the aim of many researches studies in the last decades, and that is just cause its influence in many different industries such as banking, shipping, commerce, communications, marketing, license plate recognition, etc. Due to the great importance and promising future of this field, we found useful to create a system able to recognize machine printed characters. Although, we understand that creating a system with a one hundred percent correct recognition rate is quite probably impossible in our world of noise and different font styles, because algorithms generally operate on a different set of features than humans for computational reasons. But we still can develop a system able to recognize a considerable amount of samples.Therefore, this dissertation presents a system that using feature character extraction and neural network classifier trained with the Back-propagation algorithm, can recognize machine printed English characters.The system first applies binarization to the image and avoids thinning (and other major preprocessing) by assuming that the input data is not particularly aberrant. With this assumption, it then proceeds to look for feature points.We have described two methods for feature point extraction:the first one is called discrete feature point extraction, this method scans the images looking for certain pre-defined points; the second method is zoning, which works dividing the image in zones and calculates the average gray level for each. Afterwards, the data collected from the feature extraction process is input into the neural network. This network was design with three layers and works using back propagation. Additionally, some adjustments have been made to the network in order to obtain optimum results. Finally, we performed further neural network analysis on sample data and compare the obtained results from the two proposed methods and the experiments corroborate the good the efficiency and performance of the proposed recognition system. In the last section we discuss the conclusions and future work.