This paper presents a robust approach for the automatic classification of sonar pictures. The classification task deals with the segmentation of the sea-bottom thanks to the observations given by a high resolution sonar antenna. The originality of our approach consists in describing the seabed features according to statistic properties of multiresolution parameters. The obtained statistic parameters form a feature vector corresponding to a scale parameter description. A discriminant analysis allows to significatively decrease the size of the vectorial space corresponding to the feature vectors and to generate an optimal subspace where the classification task will be done. This method has been validated on real world sonar pictures, with strong speckle noise (A training set of 300 sonar observations has been used).
This paper propose an original approach to solve the segmentation of high resolution sonar images, according to their textures. We propose a parametric method, with rotation invariance property, to classify different kind of seabeds from sides-scan sonar images. This approach has for originality to use together a discriminant analysis on the feature parameters and a multiscale analysis on the sonar images. This method is supervised in the sense that it needs a training set, allowing to generate an optimal subspace from the original feature space, by favoring decorrelated parameters. Each sonar image is described by a set of feature parameters called feature vector. These parameters are extracted from a multiresolution analysis on a wavelet basis. The main difficulty in this sonar image segmentation problem approach lies in the presence of a strong speckle noise [13][7][12]. The paper will be organized as follows. In the following section, major features of the high resolution sonar images are presented. In the same part, a compilation of the main approaches used for texture classification is briefly pointed out.