Advancing Arabic Handwritten Digit Recognition with AI-Enhanced Neural Network Architectures

Abstract

Neural network model developed in this paper aims at classification of the hand written digits using the data set from Arabic Handwritten Digits Dataset (AHDD). It also includes data preprocessing, model design, training, validating, hyperparameter optimisation, and comparison methodologies of the project. Some preprocessing included scaling of pixel intensity and data augmentation to improve variation, as well as data separation between training and validation. proposed architecture of the model were updated through adding of dropout layers as a form of regularization, tuning of the quantity of hidden layers and neurons in them, and providing dynamic form of learning rates in attempt to diminish overfitting and to improve the model’s predictive ability. The improvements obtained in classification accuracy and in sparsity of the weights of the neural net allows to underline its accuracy in recognizing the patterns of a large data set when compared to the traditional approach. However, in this study, to better assess the performances of the developed model on the AHDD, it is compared to a model built by Tariq Rashid using a raw MNIST database and various tests are conducted to point out the peculiarities of Arabic handwritten digit recognition. The study also finds avenues to improve the model beyond what is presented in this paper: 1) incorporating Convolutional Neural Network (CNN) to learn spatial hierarchies; 2) using Transfer Learning and fine tuning from pretrained models; 3) having a larger dataset which cover other patterns that may not have been included in this study. The results of this research call attention to hyperparameter optimization and architectural improvements for AI approaches to accurate digit recognition of handwritten numbers. Apart from enriching the Arabic handwriting recognition research area, this study also opens avenues for further work that seek to take these methodologies to other complex script recognitional problems in the future.