Deep Learning: segmentation of documents from the Archivo General de Indias with DhSegment and NeuralLineSegmenter

Abstract

The amount of information stored in the form of historical documents is enormous and their treatment is highly tedious. This work is intended to go one step further to facilitate the extraction of information from these documents. This is not easy since many of the historical documents are in bad condition, or their letter is practically illegible to the human eye. The aim of this project is to apply the technique of machine learning, specifically deep learning, to segment digitized images of these documents. That is, differentiate and separate the different areas that make up the document such as text, background or ornaments zones. This will allow each area to be processed separately, which would help to extract the information.