Document Image Binarization with Fully Convolutional Neural Networks

摘要

Binarization of degraded historical manuscript images is an importantpre-processing step for many document processing tasks. We formulatebinarization as a pixel classification learning task and apply a novel FullyConvolutional Network (FCN) architecture that operates at multiple imagescales, including full resolution. The FCN is trained to optimize a continuousversion of the Pseudo F-measure metric and an ensemble of FCNs outperform thecompetition winners on 4 of 7 DIBCO competitions. This same binarizationtechnique can also be applied to different domains such as Palm LeafManuscripts with good performance. We analyze the performance of the proposedmodel w.r.t. the architectural hyperparameters, size and diversity of trainingdata, and the input features chosen.