Mathematical models and neural networks for the description and the correction of typical distortions of historical manuscripts

Historical manuscripts are very often degraded by the seeping or transparency of the ink from the page opposite side. Suppressing the interfering text can be of great aid to philologists and paleographers who aim at interpreting the primary text, and nowadays also for the automatic analysis of the text. We formerly proposed a data model, which approximately describes this damage, to generate an artificial training set able to teach a shallow neural network how to classify pixels in clean or corrupted. This NN has proved to be effective in classifying manuscripts where the degradation can be also widely variable. In this paper, we modify the architecture of the NN to better account for ink saturation in text overlay areas, by including a specific class for these pixels. From the experiments, the improvement of the classification and then the restoration is significant.