A Full-Text Search System for Images of Hand-Written Cursive Documents

摘要

We propose a full-text search technique for image-scanned documents that does not recognize individual characters. The system is as fast as a full-text search of machine-readable documents. Such a system is important when working with historical handwritten manuscripts. The proposed method works independently of differences in language and font because it uses a new pseudo-coding scheme based on the statistical features of character shapes. We evaluated our method in recall-precision curves for n-gram-based query strings in Japanese manuscripts and word-based query strings in English manuscripts using two types of image features and two different pseudo-coding schemes. Results demonstrate that the precision reached over 50% at a recall point of 80% for 3-gram queries in the Japanese manuscripts. Results also indicate that our pseudo-code is suitable for applications that use machine-learning techniques. The combination of an HMM-based filtering method and our pseudo-code can significantly improve performance in terms of retrieval precision.