Real-Time Handwritten Digit Recognition using CNNs and Web Interface

Abstract

Handwritten digit recognition plays a vital role in several automation-based applications, including postal code sorting, digital form processing, and banking systems. This study presents a real-time digit recognition framework that uses a Convolutional Neural Network (CNN) combined with an interactive web interface to classify digits drawn or uploaded by users. The CNN is trained on the MNIST dataset and enhanced using preprocessing and normalization techniques to improve recognition accuracy across diverse handwriting styles. A lightweight Flask backend processes user inputs, performs inference, and returns instant predictions to the web interface. The system demonstrates efficient real-time classification and high accuracy, highlighting its potential for deployment in educational tools, automated document digitization, and user-interactive applications.

Introduction

The text describes a real-time handwritten digit recognition system (RT-HDR) that uses Convolutional Neural Networks (CNNs) to accurately identify digits written in different styles. Handwritten digit recognition is challenging due to variations in writing patterns, shapes, and noise, which made traditional methods unreliable.

Earlier approaches relied on handcrafted features and machine learning models like SVM and KNN, but these struggled with real-world data. With the introduction of CNNs, feature extraction became automatic and more effective, significantly improving accuracy.

The proposed system integrates a trained CNN model with a web-based interface, allowing users to draw digits or upload images and receive instant predictions. It is designed as a scalable, low-latency platform with components such as a user interface, backend services, deep learning inference engine, and cloud-based deployment.

The model is trained on the MNIST dataset with preprocessing techniques like normalization and data augmentation to improve robustness. The system uses REST APIs for communication and is deployed using modern technologies like Docker and CI/CD pipelines for scalability and reliability.

Experimental results show that the CNN model achieves very high accuracy (around 99.5%), outperforming traditional methods. Overall, the system successfully bridges the gap between theoretical models and practical applications by providing an accurate, fast, and user-friendly digit recognition solution.

Conclusion

This paper introduced the Real-Time Handwritten Digit Recognition (RT-HDR) system, a platform that successfully integrates a high-performance Convolutional Neural Network (CNN) into an accessible, low-latency web environment. Unlike existing research that often focuses solely on model accuracy in offline testing, the RT-HDR platform addresses the critical need for a practical, user-centric diagnostic utility by providing instantaneous feedback. The experimental results validated the platform\'s potential, highlighting the CNN\'s superior accuracy (99.5%) on the MNIST dataset compared to traditional machine learning models. More critically, the system achieved a low prediction latency (approx. 55 ms), confirming its viability for real-time application. These findings assert that the true value of deep learning in this domain lies not just in creating accurate models, but in integrating them into practical workflows that can democratize access to timely and reliable recognition services. To build upon this work, several future research directions are proposed: Expanded Character Sets: Extend the model\'s capability to recognize characters beyond standard digits, such as handwritten letters, symbols, or characters from other languages (e.g., Hindi, as referenced in the literature). Deployment on Edge Devices: Optimize the CNN architecture (e.g., quantization, model pruning) for deployment on edge computing devices (like mobile phones or microcontrollers) to enable inference without relying on cloud backend services. Enhanced User Feedback: Integrate advanced visualization techniques to show the user which parts of their drawing the CNN focused on (e.g., using Grad-CAM), providing an intuitive understanding of the model\'s decision-making process. Longitudinal Recognition: Explore the application of Recurrent Neural Networks (RNNs) or Transformers to analyze the temporal sequence of strokes (if input is captured via pen-tablet) for improved accuracy over static image classification. Adversarial Robustness: Conduct comprehensive testing against adversarial attacks to measure and improve the model\'s resilience against deliberately distorted inputs, ensuring security and reliability in less-controlled user environments. By successfully addressing the gap between high-accuracy model development and practical, real-time implementation, the RT-HDR platform offers a blueprint for next-generation web-enabled recognition systems where diagnostic accuracy, accessibility, and speed are seamlessly combined.

References

[1] \"Handwritten Arabic numeral recognition using deep learning neural networks.\" [2] A. Dutt and A. Dutt, \"Handwritten digit recognition using deep learning,\" International Journal of Advanced Research in Computer Engineering & Technology (IJARCET), vol. 6, no. 7, July 2017, ISSN: 2278–1323. [3] Y. Shima, Y. Nakashima, and M. Yasuda, \"Pattern augmentation for handwritten digit classification based on combination of pre-trained CNN and SVM,\" Meisei University, Hino-city, Tokyo, Japan. [4] A. Dutt and A. Dutt, \"Handwritten digit recognition using deep learning,\" International Journal of Advanced Research in Computer Engineering & Technology (IJARCET), vol. 6, no. 7, July 2017, ISSN: 2278–1323. [5] T. Makkar, Y. Kumar, and A. K. Dubey, \"Analogizing time complexity of KNN and CNN in recognizing handwritten digits,\" Amity School of Engineering and Technology, Amity University Uttar Pradesh, Noida, India.

Copyright

Copyright © 2026 N Lakshmi Deepthi, P Tej Bhargav, K Vishnu Vardhan, B Arjun Sai. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.