Neural networks for complex scene recognition: simulation of a visual system with several cortical areas

A general system for image interpretation based on neurobiological and physiological concepts is presented. All processing is performed by neural networks. The system emulates a robot with a single eye. Its brain has several cortical areas. It is able to learn a given number of objects. The architecture solves the problem of communication between low- and high-level processes in picture analysis. In a classical scheme of image understanding, it is necessary to perform a good segmentation to obtain a good interpretation. In the neural network approach, the concept of mental stages is used. Noisy pictures and partially occluded objects are satisfactorily recognized.<>