Ship ATR in High Resolution SAR Images via Convolutional Transformer

With the launch of high-resolution (HR) synthetic aperture radar (SAR) imaging satellites and the rapid development of convolutional neural networks (CNNs), ship recognition in HR SAR images has shown further improvements. Unlike ship targets in low and medium-resolution SAR images, which only possess a few pixels and present a spot-like appearance, ship targets in HR SAR images pose a larger area of pixels. However, CNN lacks the power to model dependencies between long-range features occupying large areas of pixels. A convolutional transformer (CvT) is introduced to deal with this issue. CvT integrates the local features of CNNs and the capability of capturing long-range dependencies of transformers to model both local and global dependencies for ship recognition in an efficient way. The cosine-margin loss is also applied to constraint strictly the distribution of the features to further improve the performance. Experimental results on the benchmark FUSAR-Ship dataset demonstrate the effectiveness of the proposed method for ship ATR in HR SAR images.