Some improvements of a rotation invariant autoregressive method. Application to the neural classification of noisy sonar images

This paper presents some improvements of a rotation invariant method based on AutoRegressive (AR) 2D Models to classify textures. The basic model and our improved version are applied to natural sidescan sonar images (with multiplicative noise) in order to extract a reduced set of relevant rotation invariant features which are then used to feed a MultiLayer Perceptron (MLP) for identification task. The basic method provides three AR parameters, estimated over a 3×3 pixel neighbourhood. We propose an extension of this method to a 5×5 pixel neighbourhood in order to take spatial interactions into account more efficiently. Three new features are estimated. Some analyses are conducted over these features to evaluate their interest. Classification results on four types of sidescan sonar images illustrate the efficiency of the proposed approach.