A Modular CNN-Based Framework for Real-Time Handwritten Digit Classification

Abstract

Handwritten digit classification is a core task in computer vision with applications in postal automation, banking, education, and digital governance. Although Convolutional Neural Network (CNN) models have demonstrated strong performance on benchmark datasets, their effectiveness in noisy real-world settings and interactive deployment remains constrained. This study introduces a modular CNN-based framework designed for robustness, scalability, and real-time execution. Three datasets: EMNIST, USPS, and a custom noisy collection were employed to capture handwriting variability, distortions, and inconsistent illumination. A preprocessing pipeline integrating adaptive thresholding, Gaussian filtering, and contour-based segmentation enhanced digit clarity and ensured consistent isolation, while dropout and batch normalization improved generalization and efficiency. The framework achieved accuracies exceeding 99% on EMNIST and USPS and 97.9% on the noisy dataset, outperforming several state-of-the-art approaches while retaining architectural simplicity. Deployment feasibility was demonstrated through a lightweight Tkinter-based graphical user interface (GUI), which supported real-time interaction. Usability testing with ten participants confirmed low-latency predictions, with most responses completed in under 50 ms, and high satisfaction ratings averaging above 4.5 on a five-point Likert scale across navigation, clarity, responsiveness, and overall experience. These results demonstrate that the proposed framework combines high accuracy with deployability in interactive environments. GUI integration serves as a supporting validation of real-time responsiveness and accessibility, while the CNN architecture remains the central contribution. Overall, the framework provides a scalable and efficient solution for handwritten digit recognition, suitable for operational use in domains such as banking, education, and digital governance.