Lightweight CNN-Based Alphanumeric OCR with Minimal Preprocessing

This paper presents a lightweight Convolutional Neural Network (CNN) model for alphanumeric Optical Character Recognition (OCR) using minimal preprocessing. The model classifies 62 characters (0–9, A–Z, a–z) from 64×64 grayscale images and is implemented using TensorFlow and Keras. Unlike traditional OCR systems that rely on complex segmentation and handcrafted features, this approach focuses on simplicity and efficiency while maintaining high performance. The model achieved 96.8% accuracy on a controlled hand-drawn dataset. Common misclassifications occurred between visually similar characters (e.g., ‘O’ vs ‘0’, ‘I’ vs ‘1’), highlighting opportunities for improved generalization and real-world robustness. This work demonstrates that effective OCR performance can be achieved with a compact architecture, making it suitable for lightweight and educational applications.