Local feature based word spotting in handwritten archive documents

In this paper we deal with a special case of archive handwritten text recognition when word spotting can be used effectively. We analyze the use of local feature descriptors and show that the Scale Invariant Feature Transform can be used efficiently despite the large variety of word shape, and the effects of different noises. We evaluate the performance on a database of 1638 word records segmented from an archive book and show that the proposed feature processing method can achieve over 80 % hit rate. Different parameter settings and variations of the local feature descriptor are analyzed.