Dai Yimian ; Zou Minrui ; Li Yuxuan ; Li Xiang ; Ni Kang ; Yang Jian

Abstract

Synthetic Aperture Radar (SAR) target detection has long been impeded byinherent speckle noise and the prevalence of diminutive, ambiguous targets.While deep neural networks have advanced SAR target detection, their intrinsiclow-frequency bias and static post-training weights falter with coherent noiseand preserving subtle details across heterogeneous terrains. Motivated bytraditional SAR image denoising, we propose DenoDet, a network aided byexplicit frequency domain transform to calibrate convolutional biases and paymore attention to high-frequencies, forming a natural multi-scale subspacerepresentation to detect targets from the perspective of multi-subspacedenoising. We design TransDeno, a dynamic frequency domain attention modulethat performs as a transform domain soft thresholding operation, dynamicallydenoising across subspaces by preserving salient target signals and attenuatingnoise. To adaptively adjust the granularity of subspace processing, we alsopropose a deformable group fully-connected layer (DeGroFC) that dynamicallyvaries the group conditioned on the input features. Without bells and whistles,our plug-and-play TransDeno sets state-of-the-art scores on multiple SAR targetdetection datasets. The code is available at https://github.com/GrokCV/GrokSAR.