الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 97 |  |  |
| | | from | 97 | | |
Abstract
Fast classification of a target according to radar cross section (RCS) signals is important for many applications especially in military application where there is a need of a system capable of separating enemy targets from friendly ones. This classification is one of the most active research and application area of Neural Networks (NNs). In this thesis, we describe the use of the NNs for object classification using collected RCS real data from coastal monostatic radar system which generates RCS polar plots of the moving ships. Our collected RCS polar plots for 3 ship classes with different sizes and details in frequency coverage (2- 18 GHz) with vertical and horizontal polarizations. These data are applied to 3 layered feed-forward Neural Networks and back-propagation training algorithm. This study proposes four models of three layered feed-forward Neural Network and back¬propagation training algorithm. In the first one, we feed 75 inputs to the NN which are the frequency, polarization index, 72 RCS values of each polar plot (regularly distributed sampled value each 50 over one rotation) and the mean of these 72 RCS values. It is found that this method for Ship classifier design offers excellent classification results when trained with our 3 ship classes RCS data where we obtained 100% correct classification for class A, class B, and class C. In the second one, we feed 4 inputs to the NN which are the frequency, polarization index, the aspect angle and its corresponding RCS value. It is found that this method offers least classification results when trained with the same 3 ship classes RCS data where we obtained 76% overall correct classification for class A, class B, and class C but it is more efficient in actual scenario because there is no guarantee that a full revolution of the object will be vie’Yed. In the third model we make a modification in the second one to increase the correct classification results so we introduce 6 inputs to the NN which are the same inputs of the second model besides the adjacent aspect angle and its corresponding RCS value, as a result the overall correct classification became 84% for the 3 classes. In the fourth model we introduce 8 inputs to the NN which are the frequency, polarization index, three adjacent aspect angle and its corresponding RCS values, as a result the correct classification became 87% so this model can be called for when we do not have sufficient information about the RCS of the whole parts of the object to be able to classify it.