TS2Anet: Ship detection network based on transformer
Dingye Liu
Abstract
Ship positioning is critical to the reduction of marine casualties and the minimisation of losses at sea. However, most existing object detection methods use horizontal box detection, which fails to detect the direction of the ship. In addition, since ships have a specific orientation, the target box identified by the horizontal box often contains impurities. This makes it difficult to accurately detect the ship's coordinates. To address these limitations, this paper proposes the TS2Anet model, which uses the S2Anet rotating box object detection network and the PVTv2 structure as its backbone. In addition, to reduce the attention to outlier samples and improve the experimental performance, the image pre-processing method “cutout” was employed to simulate occlusion due to clouds, and the GHM function was used. With an increase in accuracy of 11.98% on the HRSC2016 dataset, the proposed model showed excellent detection performance. Furthermore, the model performed well in the coastal region of the HRSID dataset. This demonstrates the accuracy of ship detection by the proposed model.