Robust Character Recognition For Optical And Natural Images Using Deep Learning

Character recognition is one of the most critical parts of many computer vision system. And many studies have explored character recognition as two subcategories: character recognition in optical images, character recognition natural images while this separation was to achieve high accuracy in each field separated but it needs more hardware resources to operate two models one for each area. In addition to that most researches divided each field into (digits recognition, character recognition) that add extra cost in both time and hardware resources also we found that both areas still have room for accuracy improvement. This paper tackles the problem of the two subclasses by building one robust, accurate classifier using a convolutional neural network that can recognize (characters and digits) accurately in both optical and natural scene images, the proposed model has been trained on a combination of EMNIST and Char74k data sets with a random data augmentation. The proposed model achieved 92% accuracy in EMINST compared to previous works shows that the proposed model has the highest accuracy among all the previous works based on EMNIST data set. We also tested the model on none-seen data sets (ICDAR203) and the obtained results indicate the high generality and the robustness of the classifier.