A Study on Real-Time Handwritten Digit Recognition Leveraging a Modified Convolutional Neural Network

Abstract

Handwritten digit recognition is a technique in which an identification program interprets and recognizes handwritten numeric digits and returns the corresponding recognition results. With the rapid advancement of deep learning, recognition technologies—including digit recognition, Chinese character recognition, and English character recognition—have undergone significant development. Digit recognition, in particular, has a wide range of application scenarios, such as enabling batch scoring in the education sector or facilitating automatic statement importation in the financial domain. This paper provides a detailed introduction to the fundamental concepts and key architectural components of the LeNet-5 network, and introduces the concept of dilated convolution as a means of network improvement. Using the well-established MNIST handwritten digit dataset, both the original and improved models are trained. Finally, a visual interface is employed to enable simultaneous real-time digit writing and recognition.