Recognizing using Deep Learning and Machine Learning to Handwrite Text

Traditional handwriting recognition methods have relied on manually crafted features and extensive prior knowledge, making Optical Character Recognition (OCR) systems challenging to train. However, recent advancements in deep learning have significantly improved handwriting recognition performance. With the exponential growth of handwritten data and enhanced processing capabilities, ongoing improvements in accuracy are crucial. The intricate structural details of handwritten characters can be captured by Convolutional Neural Networks (CNNs) with remarkable success, enabling automatic feature extraction essential for recognition tasks. Our study explores various CNN design configurations, including layer depth, kernel size, padding, receptive field, and stride size, and dilation, to optimize handwritten digit recognition. Additionally, we evaluate different Stochastic Gradient Descent (SGD) optimization algorithms to enhance performance. While ensemble models have traditionally achieved superior recognition accuracy, our goal aims to achieve similar outcomes with a CNN architecture that is pure, reducing computational complexity and testing costs. Through extensive experimentation and fine-tuning of learning parameters, we achieved a new benchmark accuracy of 99.87% on the MNIST handwritten digit dataset.