Dealing With Highly Unbalanced Sidescan Sonar Image Datasets for Deep Learning Classification Tasks

In this paper we compare several approaches to deal with the problem of unbalanced sidescan sonar datasets for classification tasks. Specifically, we consider the case where data from one class is very limited. Methods such as downsampling the over-represented classes as well as upsampling the under-represented class are analyzed. For the latter, in order to augment the dataset several generative adversarial network (GAN) architectures are compared. We also introduce a new method of transfer learning for GANs on sidescan sonar data, which we call TransfGAN and which uses ray-traced images for pre-training. On our own sidescan sonar image dataset augmentation with synthetic images from TransfGAN increases the balanced accuracy by more than 10% while also achieving a 3% higher macro F1 score. The experiments are verified on the common MNIST dataset, which is used in a downsampled version to account for the case of limited data. Here we also can achieve a better performance when augmenting the under-represented class with data generated by a GAN.