Selection of Pre-training Datasets for Sonar Image Classification

Deep learning based computer vision models like convolutional neural networks (CNN) and Vision Transformer (ViT) are more and more applied for the automatic analysis of sonar images. Since sonar image datasets typically have a limited number of samples, transfer-learning is used to train these models. However, commonly used pre-training datasets, like ImageNet, have a large domain gap to sonar images, i.e., images in these two datasets are fundamentally different. The selection of the pre-training dataset and the related domain gap have shown to have an impact on the final performance of the model. In this work, different datasets are analysed for applying transfer-learning to deep learning models for the classification of sidescan sonar images. In addition, the study is conducted for shallow CNNs, deeper CNNs as well as ViT. We quantify the domain gap using a variational autoencoder (VAE) and the t-distributed stochastic neighbor embedding t-SNE and link these values to the classification performance of the models after fine-tuning. Our results show that while no dataset leads to an improvement of all models, the Fetal dataset works well for most investigated models, while ImageNet and its grayscaled version led to a worse performance.