Automatic analysis of radar images : implementation of an artificial neural network for classification of simulated radar range profiles

In this thesis the possibility for automatic classifications of radar images has been evaluated. The topic automatic classification has always been of interest in the radar community, and the improvements both of hardware and software during the years today presents a new field for engineers. This new field of classification opens up opportunities no one would have believed in the early days of radar. One possible approach classifying radar images is using radar range profiles. Feature extraction of the peak amplitudes and their position in the profiles is a common method. In this thesis, a classification of pure range profiles without extraction of above mentioned features was evaluated. The range profiles were obtained through software simulations. The software was used to both model ground targets and simulating the range profiles. Five target models were simulated, four were modeled and one already existed in the software. All target models in the software were simplifications of real targets. Real target simplifications are made possible through the use of common geometrical figures assembled to form a more complex, but still simplified target model. Common geometrical figures used for modelling were: spheres, elliptical cylinders, rectangular plates etc. The simulation software did not include any visual confirmation of the modeled targets. A method using sinograms and the inverse radon transform was developed. This method made it possible to visually confirm the modeled targets and prevented incorrect placement or orientation of the geometrical figures. The selected classification algorithm was an artificial neural network. A neural network is a mathematical simulation of the biological nervous system. The core of a neural network consists of hidden neurons containing threshold functions performing the calculations. The network implemented was a standard 2-layer network, with non-linear threshold functions for the neurons in the hidden layer and linear threshold functions for neurons in the output layer. The range profiles were feed to the network as inputs and the outputs were vectors with five values, one value for each target model. The parameter settings of the neural network were evaluated. The neural network classification algorithm was written in the MATLAB program language. The classification results revealed that it is possible to classify pure range profiles with and without noise. The classification rates for profiles without noise were above 70% for all target models (worst case scenario with respect to the parameter settings of the neural network). Range profiles containing noise showed for a SNR level of 15dB similar classification rates as range profiles without noise.Validerat; 20101217 (root