Maritime surveillance using instance segmentation techniques
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Instance segmentation is a fundamental computer vision problem that identifies individual objects with precise boundaries to differentiate them using an instance mask. It has applications in self-driving cars, medical imaging, robotics, and automation. The research focuses on maritime surveillance using instance segmentation techniques to monitor unauthorized marine activities like illegal smuggling, human trafficking, illegal fishing, and terrorist acts. This thesis proposes deep-learning techniques for accurate and efficient ship segmentation in maritime surveillance. The study explores ship instance segmentation's implicit information to overcome challenges like scale variation, cluttered backgrounds, reflections, low-light visibility, low-resolution images, and difficulty detecting features from small objects. It surveys existing literature on object instance technology using deep learning, reinforcement learning, transformers, marine ship detection, and segmentation, and evaluates their performance on various datasets.
Second, there are just a few publicly available marine datasets for instance segmentation tasks. To address these issues, we constructed a marine dataset named ShipInsSeg, which contains 5116 images collected from YouTube, is evaluated based on various instance segmentation techniques and performance-based accuracy and frame per second (FPS). Third, we propose a novel two-stage approach for small ship instance segmentation, which enhances accuracy and captures fine details of small ships. The enhanced Atrous Spatial Pyramid Pooling (ASPP) feature fusion module is used to perform small ship instance segmentation tasks, capturing small ships that are often missed due to weak feature extraction. Experiments using three marine datasets show that the proposed model performs better in small ship instance segmentation.
Fourth, we propose an effective single-stage technique called Multiscale Attention for Single-Stage Ship Instance Segmentation (MASSNet), which utilizes attention mechanisms to improve multiscale feature extraction across dimensions, resulting in a more refined, contextually-aware representation, thereby enhancing segmentation accuracy. The technique refines boundaries, providing greater precision and detail, and is essential for marine monitoring and safety. The technique was tested using three marine ship datasets (MariboatS, ShipSG, and ShipInsSeg).
Fifth, we developed SwinInsSeg, a transformer-based technique for ship instance segmentation, using the MariboatS and ShipInsSeg datasets. The technique uses the Swin transformer backbone to learn complex data in dynamic contexts, with a multi-kernel attention module that captures features at multiple scales using attention algorithm. The research uses advanced deep-learning strategies and introduces novel techniques for identifying and segmenting ships, with promising experimental outcomes and implications for image processing, machine learning, and deep learning fields.
Please use this identifier to cite or link to this item: