Exploiting State-of-the-Art Deep Learning Methods for Document Image Analysis

摘要

This paper provides details of our (partially award-winning) methods submitted to four competitions of ICDAR 2017. In particular, they are designed to (i) classify scripts, (ii) perform pixel-based labeling for layout analysis, (iii) identify writers, and (iv) recognize font size and types. The methods build on the current state-of-the-art in Deep Learning and have been adapted to the specific needs of the individual tasks. All methods are variants of Convolutional Neural Network (CNN) with specialized architectures, initialization, and other tricks which have been introduced in the field of deep learning within the last few years.