Document Layout Analysis with an Enhanced Object Detector

Digital images of documents contain a rich set of information. To automate their extraction, computers are programmed to analyse the content of document images. Document layout analysis is vital in that respect and can enhance the optical character recognition. The boundaries of different document regions, i.e. paragraphs, figures, or tables can be estimated using the convolutional neural networks. In this paper, we present a deep neural network that is inspired by natural scene object detectors. The network is trained and tested using the labeled samples from a large public dataset. Results demonstrate the potential of using object detectors for layout analysis. An implementation of the method will be available at: https://github.com/minouei-kl/layout-detection.