Réseau de neurones dynamique perceptif - Application à la reconnaissance de structures logiques de documents. (Dynamic and perceptive neural network applied to document logical structure recognition)

Logical structure extraction of documents remains a challenging problem due to their inherent complexityand the gap between the physical features extracted from the image and their corresponding logicalinterpretation. Most of the literature approaches propose model-driven approaches which are not genericenough to handle complex and noisy documents. They do not use intermediate interpretation steps anddo not explain the relationships between the physical blocks and the corresponding logical labels. Themain objective of this thesis is to develop a hybrid method, using both data-driven and model-drivenapproach, which is capable to learn the relationships and simulate human perception during the logicalrecognition task. We have proposed a Dynamic Perceptive Neural Network which can handle drawbacksof previous systems. Four main points have been developed:– a special network topology based on local representation where the knowledge can be integrated.The logical interpretation is unfolded along the layers of the network and a training stage isperformed to find the weights for each link;– perceptive cycles (several bottom-up and top-down processes) perform the recognition. The net-work is able to generate hypothesis, validate them and detect ambiguous patterns. The contextmanages the correction of the input features to improve the recognition rate;– an input feature clustering has been proposed to speed-up the recognition. Subsets of featuresare automatically computed and are given progressively to feed the network in order to adapt theamount of computations according to the pattern complexity;– dynamic integration in the network that make it possible to integrate the data correction infor-mation during the training stage to have more appropriate behavior during the recognition. Theimprovement uses a Time Delay Neural Network architecture to take into account the input datavariations after each perceptive cycle while the recognition step is quite similar to the static one.