HANDWRITTEN DIGIT RECOGNITION USING MULTILAYER PERCEPTRON AND BACKPROPAGATION

Abstract

Handwriting recognition, a longstanding challenge in computer applications, has witnessed significant advancements with the advent of neural networks. These versatile networks find applications in various domains, including handwriting and voice recognition, as well as complex decision-making through machine learning. This project presents a Java application that employs a Neural Network to process image data, converting it into a 24 by 24 matrix on a pixel-by-pixel basis. The data is represented with values ranging from 0 to 255. The dataset is further organized, segregating it digit-wise and storing it in distinct sheets within a single Excel file located in local storage. Subsequently, the dataset is split into two distinct subsets: the training dataset and the test dataset. To optimize performance, 180 training data points are used for each digit, resulting in a total of 1800 data rows for training. The pixel values are sequentially fed into the implemented Back Propagate Neural Network Java application, developed using Java 1.8, and trained through supervised learning. During training, error is computed for the given expected output, facilitating the adjustment of the two weight matrices - one with dimensions 784 by 200, and the other with dimensions 200 by 10