Selective attention mechanisms in a vision system based on neural networks

Abstract

A system for visual recognition derived from a previously developed theoretical framework on the overall organization of the human visual system is proposed. The system operates dynamically by analyzing different parts of the input scene at variable levels of resolution through an attentional spotlight. A constant amount of information is gathered from the scene and a fixed dimension icon is produced, so that a trade-off occurs between the extension of the examined area and the level of resolution at which data are analyzed. The position of the spotlight and its dimensions are determined on the basis of the evolution of the recognition process. The icon is processed by a bottom-up path composed of a five-layer artificial neural network. The results of this net are analyzed by a planning module which determines if recognition has been achieved, or which action to undertake next. A top-down path, including a set of nets trained by the backpropagation algorithm, evaluates the parameters of the next sampling of information. The application of the system to object recognition with varying viewpoint and range from the camera is investigated.