Historical Document Text Binarization using Atrous Convolution and Multi-Scale Feature Decoder

This paper presents a segmentation-based binarization model to extract text information from the historical document using convolutional neural networks. The proposed method uses atrous convolution feature extraction to learn useful text pattern from the document without making a significant reduction on the spatial size of the image. The model then combines the extracted feature using a multi-scale decoder to construct a binary image that contains only text information from the document. We train our model using a series of DIBCO competition datasets and compare the results with the existing text binarization methods as well as a state-of-the-art object segmentation model. The experiment results on the H-DIBCO 2016 dataset show that our method has an excellent performance on the pseudo F-Score metric that surpasses the result of various existing methods.