Since sidescan sonar image data is scarce, transfer-learning of deep learning models is typically applied. The models are first pre-trained on a large dataset, e.g. ImageNet for classification or MS COCO for detection, and afterwards fine-tuned on the sonar data. Those pre-training datasets consist of natural RGB images. Sonar images, however, are grayscaled intensity images. Thus, learned features which depend on color information are useless. Using a more suited pre-training dataset could improve the classification and detection performance of deep learning models on sidescan sonar images.

In general, recent work has shown that the pre-training dataset has a strong influence when fine-tuning deep learning models, especially when the domain gap (the difference between data in the pre-training and fine-tuning dataset) is large. This is the case for sonar images and standard pre-training dataset like ImageNet or MS COCO. A survey on alternative and more suited pre-training datasets should be carried out in this master thesis. Both computer vision tasks classification and detection should be considered.

Your tasks:

• literature review on transfer-learning and the influence of the pre-training
• search for alternative datasets with smaller domain gap (e.g. from medical or radar applications)
• analyze classification and detection performance for selected datasets
• compare different deep learning architectures