Ship identification technology based on remote sensing images has made significant progress in fields such as national defense construction and fishery management. However, existing remote sensing ship research is only established in a closed environment, that is, the test set and the training set share the same categories. In actual sea conditions, ship types outside the training set will be encountered, such as new military ships. Due to the complex background of ship images, variable ship target scales, and large aspect ratios, these unknown types of ships will be directly ignored by existing models as backgrounds, which greatly affects the accurate assessment of the maritime situation. To cope with this problem, this paper proposes a network called Unknown Ship Detector (USDet). First, we design Hybrid Enhanced Attention (HEA), a module that achieves complementary global and local information processing by fusing Multi-Scale Convolutional Attention (MSCA) and Channel-Space attention (shuffle attention), enabling the model to more accurately select and emphasize the key features, which can efficiently separate foreground objects from the background. Subsequently, we incorporate the Large Selective Kernel Network(LSKNet) within the Feed-Forward Network (FFN) to adaptively adjust the receptive field for ships of varying scales. Finally, comparative experiments are conducted on the public ship data set ShipRSImageNet. Experimental results show that USDet not only effectively reduces the degree of confusion between background and unknown classes, improves the recall ability of unknown classes, but also maintains the robustness of recognition of known classes in complex backgrounds.
USDet:Unknown Ship Detection Based on Remote Sensing Optical Images
2024-11-08
9147507 byte
Conference paper
Electronic Resource
English
Multi-Visual Features Based Ship Detection in Remote Sensing Images
Trans Tech Publications | 2014
|Marine Vision-Based Situational Automatic Ship Detection Using Remote Sensing Images
Springer Verlag | 2024
|