Domain adaptation method for semantic segmentation in marine vessel inspection

The interior of the marine vessel’s ballast tank changes drastically during its operational period in harsh marine environments. An accurate semantic understanding of the ballast tank is crucial to successfully performing automated inspections. Therefore there is a need for a solution, which will guarantee good segmentation results, no matter the condition of the ballast tank. In this work we have developed a domain adaptation method, to be employed together with a segmentation model to improve the performance across different domains, to address the visually continuous degradation of the environment. The approach was to develop a photo-realistic simulation to collect synthetic data for the training of deep learning models. For the purpose of segmentation two popular architectures - ResNet and Unet were employed in various complexity variants. Domain adaptation was performed by learning domain invariant features, using a gradient reversal layer-based approach. The developed models presented an increase of 12-15% over baseline methods. The domain adaptation method significantly improved the segmentation performance of the models on the target domain, while not decreasing the computational efficiency of the models.