Applicability analysis of attention U-Nets over vanilla variants for automated ship detection

Accurate and efficient detection of ships from aerial images is an intriguing and difficult task of extreme societal importance due to their implication and association with maritime infractions, and other suspicious actions. Having an automated system with the required capabilities indicates a substantial reduction in the related man-hours of characterization and the overall underlying processes. With the advent of various image processing techniques and advancements in the field of machine learning and deep learning, specialized methodologies can be created for the said task. An intuition for the enhancement of existing methodologies would be a study on attention-based cognition and the development of improved neural architectures with the available attention modules. This paper offers a novel study and empirical analysis of the utility of various attention modules with U-Net and other subsidiary architectures as a backbone for the task of computationally efficient and accurate ship detection. The best performing models are depicted and explained thoroughly, while considering their temporal performance.