Tsinghua Bamboo Slip Scribe Verification Using Siamese Networks

Abstract The exploration of bamboo slip manuscripts and their historical and cultural significance has become increasingly challenging. Here, we propose a deep learning approach for verifying whether the scribe who transcribed bamboo slip is the same or not, exploring the handwriting style from an image processing perspective. This approach opens a novel research direction for discriminating against the scribes of bamboo slips. We focus on collecting Warring States bamboo slips housed at Tsinghua University and constructing a self-compiled dataset for scribe verification. The discrimination model resorts to an improved Siamese network framework, utilizing the enhanced lightweight MobileNetV3+ network for feature extraction and the Squeeze-and-Excitation attention mechanism for weight control. The proposed model achieves rapid and accurate verification among bamboo slip scribes. In addition, the paper elaborates on data augmentation using gray flip to increase the number of tail samples and balance between positive and negative sample pairs. The experimental results on the collected dataset indicate that the scribe verification accuracy of the proposed model reaches 90. 2%, with the area under the Receiver Operating Characteristic curve measuring 0. 96. We further validate the effectiveness of the proposed model on datasets with unclear scribe attribution.