Foreground-Background Separation by Feed-forward Neural Networks in Old Manuscripts

Artificial Neural Networks (ANNs) are widely used techniques in image processing and pattern recognition. Despite of their power in classification tasks, for pattern recognition, they show limited applicability in the earlier stages such as the foreground-background separation (FBS). In this paper a novel FBS technique based on ANN is applied on old documents with a variety of degradations. The idea is to train the ANN on a set of pairs of original images and their respective ideal black and white ones relying on global and local information. We ran several experiments on benchmark and synthetic data and we obtained better results than state-of-the art methods.