MultKMAWE: A Multi Key-Points Based Word Extraction Algorithm for Manchu Archives

Download PDF
Open Webpage

Jianjun He, Xiaoxuan Ma, Tian Tian, Yu Zhou, Wenpeng Liu, Ruirui Zheng

Published: 01 Jan 2025, Last Modified: 16 Nov 2025ICDAR (5) 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: The Manchu archives are historical records formed during the Qing Dynasty in China holding significant value, and word extraction is a crucial step in their intelligent preservation. However, the substantial variation in the lengths of Manchu words means that one-stage key-point object detection models, which have achieved SOTA performance in other ancient archives, do not perform well on Manchu archives. To address this, this paper proposes a Manchu word extraction algorithm called MultKMAWE. Using the Hourglass Network as its backbone, in addition to the five channels used in the center key-point model, the algorithm introduces three additional channels in the head module to predict the heatmap and offset of the bottom-right key-point. Then, by matching the center key-point with the bottom-right key-point, word extraction boxes are generated, effectively mitigating the impact of differences in word lengths. Additionally, based on the Qing Dynasty Imperial Household Department Archives held by the Dalian Library, this paper constructs two new word extraction datasets: the Manchu Archives Word Extraction Real Dataset (MWZS_2000) and the Manchu Archives Word Extraction Synthetic Dataset (MWTQ_2000), each containing 2,000 images. These are currently the largest datasets for Manchu word extraction. Experiments on the Manchu archives dataset (MWZS_2000) show that the proposed algorithm improves word extraction performance by 2% compared to traditional methods.