Unsupervised knowledge discovery of seabed types using competitive neural network: Application to sidescan sonar images

The conventional approaches for habitats mapping based on supervised algorithms need a seabed ground truth classes to know the entire seabed types before the training phase. These approaches give satisfying results only when a comprehensive training set is available. If the training set lacks a particular kind of seabed, it will be unknown for the classifier and the classification will be reduced to the closest known sediment class. In addition, it is not always feasible to have a ground truth samples and generally costs are very important. This is what, automated sonar systems classification are becoming widely used. This paper is concerned with automated discovery of seabed types in sonar images. A novel unsupervised approach based on competitive artificial neural network (CANN) for sidescan sonar images segmentation is proposed. The main idea is to create an unsupervised color table which allows linking between the class color and the physical nature of the seabed. This process is based on these steps. The first one consists on texture features extraction from sonar images. Secondly, Self-Organizing features maps (SOFM) algorithm is used to project the vector features on two dimensional map. Then principal component analysis (PCA) is applied to reduce the dimensionality of the result of SOFM map to only three components. The three axes obtained by PCA process will be present the RGB color table. The final result of the color table can be used for supervised or unsupervised classification of sidescan sonar images.