SAR Ship Detection Based on ViT with Quadrangle Attention and Discrete Wavelet Transform
Ziwen Wang, Jinlong Yang, Jianjun Liu · 2024
Synthetic Aperture Radar (SAR), with its capability for all-weather operation, plays an indispensable role in the monitoring of maritime vessels. Window-based attention mechanisms (e.g., Swin Transformer) demonstrate inherent limitations in addressing the challenges posed by ship targets of varying sizes and orientations. To address these issues, we construct a Vision Transformer (ViT) with quadrangle attention (QA) and discrete wavelet transform (DWT) named QD-ViT. First, we extend the window-based attention to a general quadrangle formulation, which fundamentally changes the mechanism of self-attention computation. Additionally, deformable convolution block is incorporated as the stem architecture before the attention modules to enhance the capability of the network in processing high-resolution SAR images. Finally, we add DWT down sampling modules to tackle the inadequacy of ViTs in capturing high-frequency information. Through extensive experiments on the HRSID dataset, the effectiveness of our method has been validated.