Two-Stage Deep Learning Framework for Automatic Evaluation of Hard-Pen Chinese Calligraphy

By: Wei Shi¹, Yuran Zhang¹, Yuankui Ma¹, Yu Lei², Yujie Zhao²

¹School of Science Xi’an Technological University Xi’an, China

²School of Chinese Calligraphy Xi’an Technological University Xi’an, China

Received: 2026-07-04Revised: 2026-07-29Accepted: 2026-08-07Published: 2026-09-22

IJANMC 2026, 11(4), 130-139; https://doi.org/10.58244/ijanmc.260010

The authors would like to express their sincere gratitude to the developers and maintainers of the CASIA-HWDB dataset for making large-scale Chinese handwriting data publicly available. Their contributions have greatly facilitated research in handwriting analysis and evaluation. Funding: This research was funded by the College Students’ Innovative Entrepreneurial Training Plan Program (Nos. S202510702129 and X202510702200).

Two-Stage Deep Learning Framework for Automatic Evaluation of Hard-Pen Chinese Calligraphy

Abstract—Intelligent evaluation of hard-pen Chinese calligraphy is crucial for smart education, online learning, and automated grading systems. Traditional manual assessment is inefficient, subjective, and difficult to scale. A novel two-stage deep learning framework for automatic handwriting quality assessment is proposed in this study. In the first stage, a convolutional AutoEncoder is employed to generate pseudo-labels based on reconstruction error, enabling unsupervised evaluation without manual annotation. In the second stage, supervised regression networks, including ResNet18 and MobileNetV2, are trained to predict handwriting scores from images, capturing complex structural features. It is demonstrated through extensive experiments on the CASIA-HWDB dataset that stable and reasonable score distributions are achieved by the proposed method, with predicted scores predominantly concentrated in the 650–700 range. Furthermore, comparative analysis between ResNet18 and MobileNetV2 illustrates the trade-off between model accuracy and computational efficiency, highlighting the feasibility of deploying lightweight models in resource-constrained scenarios. Overall, an effective and scalable solution for automated calligraphy evaluation is provided by the proposed framework.

Keywords-Deep Learning; Autoencoder; Resnet18; Mobilenetv2; Chinese Calligraphy Evaluation; Pseudo-Label; Two-Stage Assessment