Convolutional Vision Transformer for Handwritten Digit Recognition

Abstract

Abstract Handwritten digit recognition is an essential step in understanding handwritten documents. The state-of-the-art convolutional neural networks (CNN) methods are mostly preferred for recognizing handwritten digits. Though the recognition accuracy is high, CNN filter weights don’t change even after training. Hence the process cannot adjust dynamically to changes in input. Recently the interest of researchers of computer vision has been on Vision Transformers (ViTs) and Multilayer Perceptrons (MLPs). The use of transformer architecture enabled substantial parallelization and translation quality improvement. The inadequacies of CNNs sparked a hybrid model revolution, which combines the best of both disciplines. This paper is written to view the impact of the hybrid model on handwritten digit recognition. The testing is done on the available benchmark datasets, the Extended Modified National institute of standards and technology (EMNIST) digits dataset, and the most significant historical handwritten digit dataset (DIDA). The 10-fold cross-validation accuracy achieved on EMNIST and DIDA is 99.89% and 99.73%, respectively. The results show that the proposed method achieves the highest accuracy compared to existing methodologies. The proposed method is robust, feasible, and effective on clean and uncleaned images.