Deep Layout Extraction Applied to Historical Postcards.

We describe an experimental study on the layout extraction problem applied to circulated old postcards. This type of historical documents presents many challenging aspects related with their automatic analysis as images. For example, their degradation due to passing of time or the possible overlapping of different elements in a reduced space. Postcard layout extraction consists in segmenting in regions the various contained information types present on these images. For the proposed task, we have used semantic segmentation deep neural networks which learn to classify the document image pixels into the different considered class categories in postcards (e.g., stamps, postmarks, handwritten text or illustrations, among others). Our experiments on an annotated dataset of 100 postcards produced respective global F1-score, Jaccard and pixel accuracy metrics values of 0.92, 0.85 and 0.92, which endorses the feasibility of the proposed method. Additionally, to the best of our knowledge, this paper is one of the first investigation in this problem applied to historical postcards.