Word Spotting Application in Historical Mongolian Document Images

Abstract

This paper proposes a framework based on the word spotting technology for indexing and retrieving the historical Mongolian document images. In the framework, the scanned document images are segmented into word images by some preprocessing steps such as binarization, connected component analysis and so on. And then each word image is processed by the following procedure, including removing inflectional suffixes, feature extraction and fixed-length representation. Finally, each word image is represented by a fixed-length feature vector and considered as an indexing term. At the retrieval stage, the necessary query keyword image can be obtained by synthesizing a sequence of glyphs according to the spelling rules of Mongolian language. For word matching, the query keyword image is also converted into a fixed-length feature vector through the same procedure. And a ranking list can be returned in descending order of similarities between the query keyword image and each candidate word image. Experimental results on the data set prove the feasibility and effectiveness of the proposed framework.