Recognition of multifont English electronic prescribing based on convolution neural network algorithm

Abstract The printed character recognition is an efficient and automatic method for inputting information to a computer nowadays that is used to translate the printed or handwritten images into an editable and readable text file. This paper aims to recognize a multifont and multisize of the English language printed word for a smart pharmacy purpose. The recognition system has been based on a convolution neural network (CNN) approach where line, word, and character are separately corrected, and then each of the separated characters is fed into the CNN algorithm for recognition purposes. The OpenCV open-source library has been used for preprocessing, which can segment English characters accurately and efficiently, and for recognition, the Keras library with the backend of TensorFlow has been used. The training and testing data sets have been designed to include 23 different fonts with six different sizes. The CNN algorithm achieves the highest accuracy of 96.6% comparing to the other state-of-the-art machine learning methods. The higher classification accuracy of the CNN approach shows that this type of algorithm is ideal for the English language printed word recognition. The highest error rate after testing the system using English electronic prescribing written with all proposed font-types is 0.23% in Georgia font.