Distilling Structure from Imagery:Graph-based Models for the Interpretation of Document Images

From its early stages, the community of Pattern Recognition and Computer Vision has considered importance leveraging structural information when understanding images. Usually, graphs have been proposed as a suitable model to represent this kind due their flexibility representational power able codify both, components, objects, or entities pairwise relationship. Even though successfully applied huge variety tasks, result symbolic relational nature, always suffered from some limitations compared statistical approaches. Indeed, trivial mathematical operations do not an equivalence in graph domain. For instance, core many pattern recognition applications, there is need compare two objects. This operation, which considering feature vectors defined ℝn, properly for graphs. In thesis, we investigated perspectives, traditional graph-based methods new advances on Geometric Deep Learning. On one hand, explore problem defining representation how deal with it large scale noisy scenario. other Graph Neural Networks are first redefine Edit Distance methodologies metric learning problem, second, apply them real use case scenario detection repetitive patterns define tables invoice documents. As experimental framework, validated different methodological contributions domain Document Image Analysis Recognition.