Abstract by Christopher Tensmeyer
Context-Aware Binarization of Historical Manuscripts with Fully Convolutional Networks
Binarization of digital historical manuscript images is a common preprocessing step in a document image analysis pipeline because it greatly improves the performance on downstream processing (e.g. layout segmentation, Optical Character Recognition). However, binarization is a challenging problem due to the degraded nature of many old manuscripts. We pose binarization as a learning problem using the Fully Convolutional Network (FCN) as our model of choice. Contrary to previous learning approaches for this problem, the FCN classifies based on the spatial arrangement of pixels rather than just local image statistics. We also directly optimize a continuous adaptation of the Pseudo F-measure metric, which is commonly used to evaluate binarization algorithms. We show that our approach is flexible enough to perform well on diverse domains, such as paper documents and on palm leaf manuscripts.