Object classification using neural networks in sonar imagery
Abstract
An experimental comparison of pattern classification methods in the particular case of objects lying on the seafloor in sonar imagery is carried out. The object identification technique relies on the analysis of the object cast shadow. Different kinds of geometric features are extracted such as elongation and orientation of the shadow, Fourier descriptors, and new parameters derived from the shadow profile. The performance is evaluated using two sets of data coming form synthetic sonar images differently noised. The comparison shows better performance of multi-layer perceptron especially for poorly segmented images. Finally, the performance of the system is investigated on real images.