Contributions à la localisation adaptative de zones informatives sur des images de documents
Maroua Hammami · HAL (Le Centre pour la Communication Scientifique Directe) · 2016
Our contributions in this thesis are dealing with field spotting in colored document images. Our goal is to end up with a solution that automatically returns the right position of a region of interest (ROI), defined by a user on a reference document, into a flow of documents from the same class. As documents are provided from different and unknown sources, the position of the required information seems to be variable from an instance to another. Hence, absolute positions are considered as weak features to fill this kind of tasks. In this thesis, we propose a system that automatically localizes the information based on a relative position built with an adjacency graph. Our solution is divided into 3 modules: the first one is dedicated to turning the image into an adjacency graph built with immutable informative zones. This structure, independent of coordinates, is used to describe the position of the ROI as well as the structure of the target document. Our second contribution is related to the subgraph isomorphism tolerant to topology distortions. Our goal is to operationnalize the structure representation proposed above in order to get the best matching between the graph describing the ROI and a subgraph from the target graph describing the document layout. In the last module, we focus on an optimization problem and we propose a solution leading to evolve performances of our system by optimizing the parameters of the process line. Our technique is based on an evolutionary system and is able to be automatically adapted to the processed document class.