A Dynamically Trained Neural Network for Machine Vision Applications.pdfVIP

A Dynamically Trained Neural Network for Machine Vision Applications Nikolaos Doulamis, Anastasios Doulamis and Stefanos Kollias National Technical University of A thens Electrical and Computer Engineering Zografou 15773, Athens, Greece Tel: ++301 7722491 Fax: ++301 7722492 e-mail: ndoulam@image.ntua.gr Abstract. A dynamically trained neural network is prop osed in this pap er prop er for adapting the network p erformance to non stationary image or video inputs. The scheme includes, on one hand, a retrieval mechanism which selects the most appropriate network from the system memory and, on the other hand, a weight p erturbation pro cedure which adapts the network weights to the current condition. If no suitable network exists in memory, new weights and network structure are created and then stored for future use. Exp erimental results are provided indicating the go o d p erformance of the prop osed system to computer and machine vision applications. Key words: Neural Networks, Image Pro cessing, Adaptive Training, Non Stationary Inputs. 1. Intro duction Image understanding and computer vision, which aim at constructing meaningful de- scriptions of physical ob jects from images or image sequences, have b ecome a topic of extensive research in the recent years. The goal of these algorithms is to endow comput- ers with intelligent image pro cessing capabilities, comparable to those of human b eings. Intelligent image pro cessing enables a great deal of p otential applications. Industrial rob ots can b enet from the ability to