Deep Learning-based Object Detection Models applied to Document Images

In Document Image Analysis (DIA), which deals with solutions to obtain computer-readable description from document images, understanding and recognition of a wide spectrum of complex document images from business and financial documents to floor plans pose a key challenge due to high-level semantic information carried in such documents. The primary task is then to isolate different present contents in the documents (e.g., graphical and textual components). In this thesis, the main objective is the recognition and understanding of graphical documents in order to generate accessible graphical documents using Deep learning-based object detection models. To do so, first, the object detection in floor plans is addressed by creating and extending floor plan data sets, and then, proposing reliable detection approaches to suitably operate in real scenarios. Second, the role of transcript alignment in early printed loosely annotated texts to support word detection inside unknown images is investigated.