Page 2 of 2
A remaining problem is the lack of large-scale sonar image data sets when applying deep learning algorithms for the automatic analysis of these data. However, over the past few years, generative adversarial networks (GAN) where developed as a tool for generating synthetic data. This work investigates how GANs can be used to generate synthetic sonar images. In order to train the GAN with only a few available samples, a transfer-learning approach is applied which uses simple simulated images. Using the additional synthetic sonar images, the performance of a classifier can be increased.
side-scan sonar | autonomous underwater vehicle | deep learning | generative adversarial network | transfer-learning
