Abstract: A systematic frame for retrieving Mongolian Kanjur images by word spotting technology was designed. Some features including profile feature, projection feature and back-to-ink transition were adopted to represent Mongolian word images by deeply analyzing characteristics of handwritten Mongolian word images. By doing a lot of comparison experiments on the dataset with 5500 word images, the best combination of features was determined. Meanwhile, the Mean Average Precision (MAP) of about 78.79% was achieved and the R-Precision 73.01%. The experimental results show that the selective features are valid and effective.

Key words: historical Mongolian document image, Kanjur, word spotting, Document Image Retrieval (DIR), profile feature, Dynamic Time Warping (DTW)

摘要： 设计了一个基于word spotting技术的蒙古文《甘珠尔经》图像检索的系统框架。在充分分析了蒙古文《甘珠尔经》中手写单词图像特点的基础上，提出了采用轮廓特征、投影特征和笔划穿越数目来表示单词图像。在由5500个单词图像构成的数据集上进行对比实验，确定了最佳的特征组合，平均准确率（MAP）能达到78.79%，R-Precision能达到73.01%。实验结果表明，所选择的特征是合理的、有效的。

关键词: 蒙古文古籍图像, 甘珠尔经, word spotting, 文档图像检索, 轮廓特征, 动态时间弯曲