A Comparison of Some Morphological Filters for Improving OCR Performance

Studying discrete space representations has recently lead to the development of novel morphological operators. To date, there has been no study evaluating the performances of those novel operators with respect to a specific application. This article compares the capability of several morphological operators, both old and new, to improve OCR performance when used as preprocessing filters. We design an experiment using the Tesseract OCR engine on binary images degraded with a realistic document-dedicated noise model. We assess the performances of some morphological filters acting in complex, graph and vertex spaces, including the area filters. This experiment reveals the good overall performance of complex and graph filters. MSE measures have also been performed to evaluate the denoising capability of these filters, which again confirms the performances of both complex and graph filtering on this aspect.