Neural network feature vectors for sonar targets classifications

Abstract

This paper deals with classification of objects using active sonar and neural network classifiers. Various methods of feature extraction, i.e., transformations of the recorded data to a form better suited as inputs to the classifier, have been applied to recorded active sonar data. The well-known multilayer perceptron architecture has been used. To reduce the computational effort and the number of parameters to estimate, the number of input nodes were limited, and it has been shown that feature vectors using Walsh transform coefficients are superior to feature vectors using Fourier transform coefficients under this constraint. Particularly with low SNR, the feature vectors based on Walsh transform coefficients outperform the Fourier based vectors. Some minor effort has been expended on optimizing the neural network implementation.