Ship Deck Segmentation Using an Attention Based Generative Adversarial Network

In this paper, we proposed an attention-based generative adversarial network (GAN) model to clean low quality, noisy engineering documents of ship decks provided by the Military Sealift Command (MSC). We proposed four loss functions and optimized the hyperparameters associated with the model and found that each of the parameters had significant influence on the model performance. We trained the proposed model with 154 MSC noisy/clean image pairs with different parameter combinations and applied the trained model to 9 MSC images to identify the best parameter combination. Experiment results show that the proposed model could achieve good results with the limited dataset and the cleaned documents can be used for 3D modeling of ships that will revolutionize ship maintenance.