Comparison of Machine Learning Models for Multiple Hand-Written Digits Recognition

Hand-written digit recognition is the process of enabling machines to identify human hand-written digits. A highly proficient algorithm is required when dealing with handwriting recognition due to its complexity. A neural network is a mathematical representation of how the brain functions. It comprises numerous layers with numerous activations, these activations mimic the neurons in the human brain. To recognize the underlying relationships, a neural network attempts to acquire and learn a set of parameters from a set of data. Neural networks yields better results as they have the ability to adapt to changing input, without changing the output criteria. The MNIST dataset is widely used for the hand-written digit recognition process and it has 70000 handwritten digits. In this paper, convolution neural networks are used to train these images and build a machine-learning model. In order to reduce the illumination, image binarization and reshaping are carried out on the images. The model is successfully trained and reached a validation accuracy of 99.27%. To make it user-friendly canvas is created with the help of GUI where the user can write multiple hand-written digits and in return the image is analyzed by the model and the detected result can be viewed in the canvas. The processing of bank checks, filling out forms, and sorting mail are a few uses for digit recognition. CNNs are usually regarded as the best algorithm for hand-written digit recognition with a test accuracy of 99.26%, followed by SVM with a test accuracy of 89.34% and LeNet-5 with a test accuracy of 97.13%.