ABSTRACT
To deal with the huge amount of digitized documents, an increasing interest is accorded for developing retrieval systems, which aims to index or extract information from handwritten data. Specifically, Writer Retrieval systems exploit the writer profile to find all its documents available in a given database. Retrieval is carried out using document pages that undergo feature generation before being introduced to the matching step. This work investigates the performance of different kinds of feature generation schemes such as local binary pattern (LBP), gradient local binary pattern (GLBP), histogram of oriented gradient (HOG), run length feature (RLF), and pixel density. Various similarity and dissimilarity measures are used to achieve document matching. Experiments are conducted on an ICDAR-2011 database that contains handwritten documents of 26 individuals. The results obtained evince that RLF associated with the Manhattan distance outperforms all other systems as well as the state of the art.
