A recognition model for handwritten Persian/Arabic numbers based on optimized deep convolutional neural network

Recently, artificial intelligence-based applications are universally acknowledged. Digit recognition, particularly Persian/Arabic handwritten digits, has many applications in today's commercial contexts for example office automation and document processing. However, researcher are struggling in hand-crafted digit scripts due to the presence of different digit writing patterns, cursive nature and lack of large public databases that make the feature extraction process more complex. Therefore, critical investigation is needed to reduce these challenges. In this study, a modified Deep Convolutional Neural Network (DCNN) architecture using three convolutional layers blocks based on convolution, batch normalization, pooling, fully connected and dropout regularization parameters are employed to hinder overfitting and increase generalization performance is proposed to recognize handwritten digits. Initially, digits taken from the HODA database are pre-processed using various steps including smoothing, black and white images to grayscale intensity images conversion and resizing it to a fixed dimension. Then optimal features extraction and recognition of handwritten images are done by the DCNN algorithm. In deep learning domain, optimization algorithms are considered core solution and their performances highly depends on optimization algorithm selection. In this paper, various optimization algorithms such as stochastic gradient descent (SGD), Adam, Adadelta, Adagrad, Adamax, Momentum, RMSprop and Nag are employed for the optimization of proposed DCNN. Moreover, current research also analyzes the role of different epochs to ameliorate optical character recognition (OCR) performance of Persian/Arabic handwritten digits. At the end, we also worked on finding a suitable composition of learning parameters to establish high performance DCNN architecture that conquer the loopholes of traditional methods. Results reveal that the proposed DCNN model achieves state-of-the-art performance and outperform other studies in the literature.