Sea-floor classification by neural hybrid system

High-resolution side-scan sonars perform high-quality sea-bed imaging. In these images, texture recognition is a reliable characterization of seafloor nature. This task is efficiently carried out by previously trained neural classifier. Once the classifier is chosen; it is trained to separate various classes of textures according to examples extracted from homogeneous regions labeled by an operator. Most of the classifiers are not trained directly on gray-level images, instead, the learning set consists of features vectors computed images. Generalization capabilities of neural networks ensure a correct classification of test images provided that they are close the training set examples. Unfortunately, poor performances are observed should the examples slightly differ from the training set or the validity range of the features is to narrow. In this paper, the authors present a hybrid neural classifier consisting in a supervised network driven by an unsupervised one. The original image is partitioned in small two-dimensional elements from which feature vectors are extracted. The vectors are clustered by a self-organizing feature map and each element is labeled according to its activity. A supervised neural network, suited for Incremental learning, is trained on these elements. The self-organizing map searches for occurrences of new textures by continuously analyzing features extracted form sonar images. When such event arises, incremental supervised networks update its knowledge. Several improvements stem from this classification paradigm. The unsupervised network provides a quantitative hint of feature-set discrimination ability, class-selection method ensures a strong coherence between the feature set and the texture classes and, finally, incremental learning reduces the initial training phase foregoing exploitation and allows on-line incorporation of additional knowledge