Neural network as edge detector

Detection of edges in the image is extremely important task for many of today's robotic control systems. With this procedure we detect the edges of objects and boundaries between objects and background in the picture, and such obtained information we use in extraction and object segmentation from the picture. There are several ways to detect edges in the picture, but they all can be grouped in two basic groups, gradient detectors and zero crossing detectors. This paper presents an alternative to standard edge detectors in form of back propagation neural network which is trained over the results of “classic” edge detector (Sobel), and has been shown that in most cases when there is presence of noise in the image, neural network model gives significantly better results. The structure of neural network in this paper has three layers (input, hidden and output layer); it is feed-forward multilayer back propagation neural network.