A Lightweight Web Platform for Real-Time Kuzushiji Recognition via One-Shot Pruning

Kuzushiji, a classical Japanese cursive script used for more than one thousand years, is preserved in a vast number of historical documents that constitute an important part of Japan’s cultural heritage. However, the complexity and variability of Kuzushiji characters make these documents difficult to access for non-specialists, creating challenges for the digitization, preservation, and dissemination of historical knowledge. To address this issue, this study proposes the Lightweight Kuzushiji Recognition System (LKRS), a lightweight recognition framework designed for 1120-class Kuzushiji recognition and practical online deployment. LKRS integrates modern lightweight neural network architectures with a one-shot pruning framework to reduce model complexity while maintaining recognition performance. In addition, a web-based interface is developed to support interactive character selection and repeated recognition, providing convenient access for researchers and the general public. To evaluate the effectiveness of the proposed framework, extensive experiments are conducted on multiple public datasets and different backbone networks. The optimized EfficientNet-B0 model achieves a recognition accuracy of 94.08% on the Kuzushiji dataset while reducing the number of parameters and floating-point operations (FLOPs) by 62.50% and 86.91%, respectively. Experimental results on Fashion-MNIST, CIFAR-10, and Kuzushiji further demonstrate the effectiveness of the proposed framework in reducing model complexity while maintaining competitive classification performance across different datasets and network architectures. The proposed LKRS provides an efficient and accessible tool for Kuzushiji recognition and contributes to the digitization, preservation, and dissemination of Japanese cultural heritage. Furthermore, the framework offers a reusable reference for developing lightweight recognition systems for other historical scripts.