Enhanced Mask-RCNN for Ship Detection and Segmentation

Abstract

Ship detection and segmentation using satellite remote sensing imagery have become a hot issue in the scientific community. This sector helps control maritime violence, illegal fishing, and cargo transportation. Most available methods used for detecting ships perform object detection but don’t perform semantic segmentation. Besides, previous papers had various flaws, such as the inability to detect small ships, greater false positives, and severe noise in the given Synthetic Aperture Radar (SAR) images, which affects low-level feature learning in shallow layers and makes object detection more difficult. The intricacies of SAR images also significantly reduce the benefits of Convolutional Neural Networks (CNNs). Then there are a few models that can’t tell the difference between things that appear to be ships. Furthermore, some existing models performed somewhat worse than other state-of-the-art given and successful frameworks. The cost of computing resources is higher for some models. This research provides a ship detection approach based on an improved Mask Region-Based Convolutional Neural Networks (Mask RCNN). At the pixel level, the proposed approach can detect and segment ships. For object detection and segmentation, Mask RCNN comprises two parts. In the segmentation part, more Convolutional Layers are added, and hyperparameters are changed to improve the overall output. Because of these changes, the proposed model can work more accurately than existing models. Using this mentioned method on the Airbus Ship Detection dataset on Kaggle, we achieved an accuracy of 82.9% on the proposed model.