Convolutional Neural Network architectures for digit recognition system

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] Rapidly evolving of Convolutional Neural Networks (CNNs) appeals and endeavors us to explore and discover various CNN model's robustness leading to establishing more effective and efficient models achieved state-of-the-art results on challenging datasets including novel benchmark used in this work. Comparing with most contemporary works, our models outperform the accuracy demonstrated by all existing models. Enhancement performed during this work is not only included model performance but also compromise other vital deep learning aspects such as speed and efficiency. Moreover, a new challenging Arabic handwriting digit dataset is also introduced in this work. We create novel dataset because recently handwritten digit recognition becomes vital scope and it is appealing many researchers. Also the lacking research of using Arabic digits endeavors us to dig deeper by creating our challenge Arabic Handwritten Digits which consists of more than 45,000 samples.