Efficient multiscale Sauvola’s binarization

This work focuses on the most commonly used binarization method: Sauvola’s. It performs relatively well on classical documents, however, three main defects remain: the window parameter of Sauvola’s formula does not fit automatically to the contents, it is not robust to low contrasts, and it is not invariant with respect to contrast inversion. Thus, on documents such as magazines, the contents may not be retrieved correctly, which is crucial for indexing purpose. In this paper, we describe how to implement an efficient multiscale implementation of Sauvola’s algorithm in order to guarantee good binarization for both small and large objects inside a single document without adjusting manually the window size to the contents. We also describe how to implement it in an efficient way, step by step. This algorithm remains notably fast compared to the original one. For fixed parameters, text recognition rates and binarization quality are equal or better than other methods on text with low and medium x-height and are significantly improved on text with large x-height. Pixel-based accuracy and OCR evaluations are performed on more than 120 documents. Compared to awarded methods in the latest binarization contests, Sauvola’s formula does not give the best results on historical documents. On the other hand, on clean magazines, it outperforms those methods. This implementation improves the robustness of Sauvola’s algorithm by making the results almost insensible to the window size whatever the object sizes. Its properties make it usable in full document analysis toolchains.