Performance analysis for various clutter elimination methods in ship target classification

Target classification studies with radar range profiles, which are very popular in recent years, continue to be a difficult problem during classification processes due to the presence of environmental clutter factors in the measuremental data. In this dissertation, performance analysis of three different clutter elimination algorithms has been carried out in order to remove clutter from range profiles. Performance analyzes of the clutter elimination algorithms were carried out using range profiles obtained via measurements related to nine different marine targets. In order to investigate the effects of the clutter removal performance as well as the classification results, the clutter-free range profiles have been provided as an input to classifier structures and the effects on the classification performance have been demonstrated. In this context, simple thresholding which is commonly used in the literature in clutter elimination processes as well as non-local means- (NLM) and singular value decomposition- (SVD) based clutter elimination methods that have not been used in this area are considered. The performance analysis of the clutter elimination methods are compared in the manner of mean square error rate metric and outlier analysis method. In addition to this, to determine the effect of considered clutter elimination methods on target classification performance, clutter-free range profiles of nine marine targets are fed to classifier structures as inputs and the results are examined. One of the used classifier structures is the deep learning-based convolutional neural network (CNN) which has an increasing popularity in recent years. Artificial neural networks-based probabilistic neural network (PNN) is used as the other method. In the CNN method, the feature extraction is performed automatically in the classifier structure, while in the PNN classifier structure, manually extracted two dimensional wavelet coefficients are used as the input features. The purpose of selecting wavelet-based features is to investigate that whether the clutter elimination methods are sufficient alone or not, and also to find out which method should be followed. At the end of this dissertation, results of marine target classification simulations using measuremental range profiles will be given. By examining the classification results, it is discussed how the clutter elimination methods affect the classification performance and which method is effective for measuremental range profiles