Handwritten Digit Classification using Iterative Haar-like Operators and Neural Networks

Although automatic handwritten digit recognition is increasingly finding usage in mobile and hand held devices, it remains a challenging problem. Despite the incorporation of modern techniques such as deep learning, there is room for improvement in the accuracy rates and running times of existing approaches. This study proposes the use of Haar-like features and artificial neural networks to complement existing efforts in handwritten digit recognition. Haar-like filters of different granularity levels are applied to images of handwritten digits, in order to determine the effect of increased granularity on classification effectiveness. Experiments on the MNIST dataset show that increasing Haar-like filter granularity yields improved classification accuracy. Using 256 Haar-like features, a neural network is trained and tested on 60000 and 10000 handwritten digit images, respectively; achieving an overall accuracy of 89%.