Graphics Extraction from Heterogeneous Online Documents with Hierarchical Random Fields

Graphical objects are important elements of freely handwritten notes but their segmentation from the document is challenging due to their irregular properties. This paper introduces an original solution for automatically segmenting diagrams and drawings from unstructured online documents. We propose a multi-scale representation of the document modeled as a hierarchical Conditional Random Field to predict the detection of graphical elements at the stroke level. An experimental evaluation with realistic documents highlights the benefit of the hierarchical model in comparison with a flat Conditional Random Field and demonstrates the robustness of our system.