Neural Networks Applications

Abstract

This thesis describes three image recognition models. The applications on hand utilize the eigenvector technique and the neural networks for aircraft identification purpose. The methods of image identification are reviewed using statistical, syntactic, and neural network. In the first method, the contour of each aircraft image is isolated and clipped using image pre-processing operations. Image silhouettes normalization followed by feature extraction are done using the principal component analysis. A highly recognition success is obtained with x-features, where x is the number of referenced aircraft.

An aircraft identification system based on back-propagation neural network is presented. The effect of activation function and number of hidden neurons on recognition performance is studied through this work. The last approach is based on Kohonen and Grossberg models. It was found that the aircraft recognition method based on back-propagation and counter propagation approaches has been succeeded to recognize the test images with signal to noise ratio greater than -5.9 dB and -9. I I dB respectively. In addition, the above two systems have positive response when the test images became uncompleted with percent until 14% and 40% respectively. The neural network systems are tested on 1508 aircraft

images some of them are noisy and uncompleted.

The noise associated with aircraft images has been measured and found to be uncorrelated. All the presented aircraft recognition methods are invariant to translation and rotation except the first one is not sensitive to scale variation.