Header for SPIE use Neural network based on multi-valued neurons and its application to image recognition and blur recognition Igor Aizenberg * , Constantine Butakoff * company Neural Networks Technologies Ltd. (Israel) ABSTRACT Some important ideas of image recognition using neural network based on multi-valued neurons are being developed in this paper. We are going to discuss the recognition of color images, which is reduced to recognition of gray-scale images. An approach, which has been developed, is illustrated by simulation results. Recognition of distortion (blur) types, distortion parameters and recognition of images with distorted training set using the same neural network is also considered. At this time Gaussian blur and motion blur were taken as distortions. This part of work is also illustrated by simulation results. Keywords: neural network, multi-valued neuron, image recognition 1. INTRODUCTION Idea 1 of image recognition using neural networks based on multi-valued neurons has been proposed less than 3 years ago. During this short time the proposed approach has been developed 2-5 . A Multi-valued neural element (MVN), which is based on the ideas of multiple-valued threshold logic 3,6 , has been introduced in 7 . A comprehensive observation of multi-valued neurons can be found in 3 . Different kinds of networks that are based on MVN, have been proposed 3,5,8 . Successful application of these networks to simulation of the associative memory 3,5,8 , image recognition and segmentation 3,9 , time-series prediction 3 confirms their high efficiency. Highly effective quickly converged learning algorithms for MVN and neural networks based on them have been elaborated 3 . Here we are going to develop some important ideas of image recognition using MVN-based neural network. Solution of the image recognition problem using neural networks became very popular during last years. Many corresponding examples are available 2,3,5,8,10,11 . On the other hand many authors consider image recognition reduced to analysis of the orthogonal spectrum coefficients using different neural networks 3,5,11,12 . We would like to develop here an approach to the image recognition presented in 2,3 . This approach is based on the following background: 1) high functionality of multi-valued neurons and quick convergence of the learning algorithm for them; 2) well-known fact about concentration of the signal energy in the low-frequency part of orthogonal spectra 12 . Different MVN-based neural networks were already used for solution of the image recognition problem. Several papers devoted to different types of an associative memory should be mentioned. MVN-based cellular neural network has been proposed as associative memory in 7 . MVN-based neural network with random connections has been proposed as associative memory 3 alternative to the Hopfield one. The MVN-based network with random connections 3 requires much smaller number of the connections than fully connected Hopfield network 8 . The quickly converged learning algorithm is another important useful property of the MVN based network with random connections. On the other hand Hopfield-like MVN- based neural networks have been also proposed as associative memory 5,8 . We will concentrate here on the development of image recognition approach using single-layered MVN-based neural network 2-4 . This approach is based on the analysis of orthogonal spectra coefficients on the MVN-based neural network. Because of the fact that inputs are taken from Fourier spectrum, and FFT is widely used for Fourier spectrum calculation, recognition of images of any dimensions will be discussed, not only of images, which dimensions are powers of two. We are also going to discuss the recognition of color images, which is reduced to recognition of gray-scale images. * Correspondence: (IA): E-mail: igora@netvision.net.il and igora@nnt-group.com ; Mapu 18, ap. 3 , Tel Aviv, 63434, Israel (CB): E-mail: cbutakoff@nnt-group.com ; NNT Ltd., 155 Bialik str., Ramat-Gan, 52523 Israel