Using artificial neural networks for classification of black-and-white images
- PDF / 362,744 Bytes
- 4 Pages / 595 x 842 pts (A4) Page_size
- 8 Downloads / 231 Views
USING ARTIFICIAL NEURAL NETWORKS FOR CLASSIFICATION OF BLACK-AND-WHITE IMAGES
UDC 519.683.004.424
E. S. Borisov
This work is devoted to the problem of automatic classification of binary images. To solve this problem, neural networks are used. Experiments based on a database of faxgrams are performed with neural networks of different types. Keywords: neural network, image classifier. INTRODUCTION In this paper, the problem of classification of black-and-white images is considered. To solve it, a system based on the apparatus of artificial neural networks is used. Experiments are performed with neural networks of different types. For the experiments, a database of facsimile messages was used as a data set, i.e., the classification system developed processes black-and-white pictures with 1728 points along the horizontal direction and, in the general case, an unlimited number of points in the vertical direction (tapes). GENERAL SCHEME OF THE CLASSIFICATION SYSTEM The classification system consists of two component parts (Fig. 1), namely, a characteristic function and a neural network. Characteristic function associates with an input picture a point (a characteristic) in the feature space. Thus, a set of points is obtained from a collection of pictures. A characteristic must be computed in an acceptable time. The obtained characteristic vector must be informative in order that it can be efficiently processed by the neural network but, at the same time, must be small in order that this process can run at a sufficient speed. To efficiently classify pictures, it is necessary that pictures of different types be mapped by the characteristic function into different spatial domains of the feature space and that these points should not be chaotically mixed. An ideal characteristic function forms several (according to the number of given classes) nonintersecting domains in the feature space, and each of them contains characteristics of pictures of only one type. During the experiments performed, directly connected neural networks with different numbers of layers and also a Kohonen neural network were used as neural network classifiers. DIRECTLY CONNECTED NEURAL NETWORK The topology of a directly connected three-layer neural network is shown in Fig. 2. The characteristic of a picture ( x 0 ,... , x m ) is applied to its input, and a vector ( y 0 ,... , y m ) determining the number of the class to which the input image belongs is produced at its output. Thus, the size of the input layer is equal to the size of the Cybernetics Institute, National Academy of Sciences of Ukraine, Kiev, Ukraine, [email protected]. Translated from Kibernetika i Sistemnyi Analiz, No. 2, pp. 184–187, March–April 2008. Original article submitted February 15, 2007. 304
1060-0396/08/4402-0304
©
2008 Springer Science+Business Media, Inc.
Picture Characteristic function Characteristic Neural network classifier Result Layer 3 Fig. 1. General scheme of an image classification system.
Layer 0
Layer 1
Layer 2
Fig. 2. Layout of a directly connecte
Data Loading...
