Layout Analysis of Historical Document Images Using a Light Fully Convolutional Network.

In the last few years, many deep neural network architectures, especially Fully Convolutional Networks (FCN), have been proposed in the literature to perform semantic segmentation. These architectures contain many parameters and layers to obtain good results. However, for Historical document images, we show in this paper that there is no need to use so many trainable parameters. An architecture with much fewer parameters can perform better while being lighter for training than the most popular variants of FCN. To have a fair and complete comparison, qualitative and quantitative evaluations are carried out on various datasets using standard pixel-level metrics.