Detecting Fine-Grained Airplanes in SAR Images With Sparse Attention-Guided Pyramid and Class-Balanced Data Augmentation

Airplane detection in synthetic aperture radar (SAR) images has drawn much attention owing to the success of deep learning methods. However, development fine-grained airplane SAR is still a dilemma due small inter-class variance and large intra-class complex scenes with strong interference from background. In addition, class imbalance problem multi-class recognition also significantly limits direct application general deep-learning-based detectors. This paper proposes two effective methods tackle above problems, respectively. First, we propose sparse attention-guided pyramid (SA-FP) module simultaneously sample discriminative local features scattered multi-scale layers adaptively aggregate them better classify subordinate-level airplanes multiple scales. Second, simple class-balanced copy-paste data augmentation (CC-DA) strategy, which randomly copies an one category pastes it onto image according class-wise probability, proposed for balance. Finally, extensive experiments on public dataset three representative benchmarks are conducted show effectiveness generalization The combination these based Cascade R-CNN benchmark won fifth place 2021 Gaofen challenge.