CCR-LWECNN: A Lightweight CNN Framework for Chinese Calligraphy Recognition and Evaluation

Abstract

This study presents a lightweight enhanced CNN architecture (CCR-LWECNN) for Chinese calligraphy recognition, addressing the challenges of multi-class classification across 12,152 labeled images spanning 960 Chinese characters in five calligraphic styles. Unlike previous studies limited to small character sets and single recognition approaches, this research integrates character recognition with image processing techniques. Data augmentation using TensorFlow’s Image Data Generator—applying rotation and zoom—was employed to improve class balance and variety. The proposed model, comprising five convolutional and three fully connected layers, processes 224×224-pixel images and leverages pretraining for robust feature extraction. CCR-LWECNN achieved superior performance with 96.5% accuracy, 95.6% precision, 95.2% recall, and 95.6% F1-score, outperforming baseline models such as traditional CNN (90.5%), SVM (85.2%), and Random Forest (75.4%). By effectively mitigating overfitting and underfitting through dropout layers and augmentation, this approach advances automated Chinese calligraphy recognition and provides a scalable solution for real-world applications.