Using Harris Corners for the Retrieval of Graphs in Historical Manuscripts

In recent years, several methods have been proposed for content-based retrieval from manuscripts, mostly based on character or word similarity. In this paper, we present a new segmentation-free method, called Harris Corner Matching (HCM), which accepts an arbitrary writing pattern as a model and allows to retrieve similar patterns from a possibly large database. Retrieval is performed in two steps. In the first step, candidate targets are determined by comparing configurations of interest points in query and data. In fact, interest points can be used as precomputed indices. In the second step, deviations between the interest point configurations of query and candidate target are used to warp the target, this way adapting it to the query. A final evaluation is obtained by template matching after binarization of query and target. The method has been evaluated for retrieval from historical Chinese and Sanskrit manuscripts and has shown good results.