Combination of Two Fully Convolutional Neural Networks for Robust Binarization

To be able to process historical documents, it is often required to first binarize the image (background and foreground separation) before applying the processing itself. Historical documents are challenging to binarize because of the numerous degradations they suffer such as bleed-through, illuminations, background degradations or ink drops. We present in this paper a new approach to tackle this task by a combination of two neural networks. Recently, the DIBCO binarization competition has seen a growing interest in the use of supervised methods to binarize challenging images. Inspired by the winner of the DIBCO 17 competition, which uses a fully convolutional neural network (FCN), we propose a combination of two FCNs to obtain better performance. While the two FCNs have the same architecture, they are trained on different representations of the input image. The first one uses downscaled image to capture the global context and the object locations. The second one works on patches of native resolution to help defining precisely the boundaries of the characters by capturing the local context. The final prediction is obtained by combining the results of the two FCNs. We show in the experiments that this strategy provides better results and outperforms the winner of the DIBCO17 competition.