Detecção Hierárquica de Classes Desconhecidas em Sonar por "Autoencoders" Convolucionais

Acoustic waves captured by passive sonar systems are analyzed by human operators, aiming to identify possible threats in the subsea environment. Automatic Classification Systems can aid in the work of this professional, however requiring mechanisms to deal with the presence of unknown classes. This article proposes the use of a hierarchical committee of convolutional autoencoders to build these systems, as a more robust alternative to the k-nearest neighbors algorithm, which represents the state-of-the-art in this problem. Real data belonging to 8 classes of ships under different operational conditions were evaluated. Results signalize a competitive performance of the proposed technique.