Binarisation d'images de documents dégradés en utilisant l'apprentissage automatique

Abstract

Binarization is an important step in any image processing and analysis process and more particularly, document images. Indeed, a large number of techniques have been proposed in the literature for the binarization of grayscale or color images, each of which is appropriate for a particular type of images, but unfortunately none of them show effective for binarization of degraded old document images. The latter are known for their poor quality due to the various deteriorations undergone during the life cycle of the document and the provisions hitherto employed for their safeguard. However, the majority of the methods proposed in the literature are based on the calculation of one or more thresholds to perform the binarization. They are called as follows: binarization methods by thresholding. In this thesis we propose a new method of binarization of images of degraded documents which is not based on thresholding (like most binarization algorithms), but on machine learning, a multilayer Perceptron type artificial neural network. The latter is known for its ability to generalize from a limited set of so-called learning behavior data for other data that have not been learned. The proposed method is tested on a collection of public images and the results obtained are encouraging.