Disentanglement learning for Synthetic Aperture Sonar imagery

The performance of Convolutional Neural Net (CNN) based target recognition algorithms for Synthetic Aperture Sonar (SAS) imagery is strongly dependent on the seafloor background, with worse classification performance for samples with backgrounds outside the CNN training curriculum. Online performance estimation seeks to identify whether new samples had representation within the training curriculum, and if so whether a given model is expected to perform well on that sample. Performance estimation is critical in underwater object classification, given the diversity of seabed textures and sparsity of training data. Disentanglement learning is a data-driven approach to learning a latent representation that is conceptually interpretable.In the context of performance estimation, we use a disentanglement learning approach based on variational autoencoders to train a latent representation that encodes separately both target class and background type. Online performance estimation is performed by determining if a given sample has a background that is likely to be in the class of backgrounds included in the training data. Results are demonstrated using in-air linear SAS scans from the airSAS dataset.