Application of neural networks to the search of objects in an image

Yurii Yakymenko¹ and Vitalii Dzyuba¹

¹Faculty of Electronics, National Technical University of Ukraine “KPI”, 03056, 16 Polytechnichna st., room 201, Kiev, Ukraine; e-mail: dzyuba@ee.ntu-kpi.kiev.ua

The pattern recognition is one of fundamental problems, whose solution implies classification of objects on images. For identification of images various systems of technical vision are used permitting to find a certain object on the image, and to identify it. As a rule, the methods of search are aimed to solving some specific problem for a certain class of objects, so that the problem of creating an instrument aiding in seeking for various objects on images is a pressing task. The purpose of this work consists in development of a neural-network algorithm with high ability of classification — in order to seek effectively for different classes of objects on images. An integral part of the detector responsible for the search represents a neural network falling in the class of associative-type computation devices. The learning of the network, i.e. its ability to identify different objects, was carried on with the use of the AdaBoost magnification method. For the objects we took images of human faces and eyes, images of car number-plates, and of passive components on a printed circuit board. To perform teaching and testing the detector, we have developed a package of programs were written in the C++ programming language and compiled in Microsoft Visual Studio 2005. As shown by the tests, the newly developed approach to search for different classes of objects with the use of a neural network is very effective. According to the tests, the percentage of correctly identified objects of the selected images was 99.8% when checking human faces, 98.9% for human eyes, 96.2% for car number-plates, and 96.4% for passive elements on a PCB. The obtained results confirm high performance of the neural network, which turns out to be a versatile instrument adding to right and efficient finding of different classes of objects on different images.