Hierarchical neural networks for image interpretation. (Q1405985)

This booklet is the reprint of a thesis. It addresses image interpretation using a neural network architecture mimicking the human visual system. The main characteristics of the ``Hierarchical Neural-Network'' proposed by the author are: a 2D horizontal organisation of the layers, a vertical growing level of abstraction of the layers, with the level of abstraction, the spatial resolution decreases and the coded features diversity and invariance increase, connections between the layers are bottom-up, top-down and local lateral, neural weight are shared. Machine learning techniques (supervised and unsupervised) are used to perform learning. A method combining backpropagation and RPROP optimisation is presented to get fast and stable learning. The exposition is divided in two parts, namely theory and applications. After a general introduction (1) the theory part begins by an interesting survey of the neurobiological background (2) and of the related works in neural network for image understanding (3). A chapter is devoted (4) to present the proposed architecture, followed by unsupervised (5) and unsupervised (6) learning techniques used. The application part treats various applications, some real size industrial applications as meter value recognition, binarization of matrix codes as well as image reconstruction and faces localisation. The results obtained prove the efficiency of the method, one the main difficulty being to get a non sequential implementation of the whole system. In short this thesis is very interesting, well written and easy to read. If it is interesting for people working in this field, it can also provide a survey approach of the utilisation of neural network in vision application to the non specialist.