Abstract: Aiming at the problem of insufficient detection accuracy of rotated small objects in remote sensing images under  limited scale, diverse orientations, and complex background conditions, we  proposed a Transformer network  method with dual-domain query enhancement and rotation awareness. The method used  convolutional neural network  to extract multi-scale features and introduced a joint enhancement in both spatial and frequency domains at the encoding end. The  spatial  adaptation module captured geometric structure features by using multi-scale receptive fields, while a frequency adaptation  module  extracted directional information through  wavelet transform. After  cross-domain fusion, a feature query  with both spatial and frequency 
perception capabilities was generated. We introduced a rotation-aware module at the encoding end to dynamically estimate spatial offsets during the Transformer decoding process, achieving precise alignment of rotated small objects at multiple scales. The experimental results show  that the proposed method significantly improves detection accuracy of rotated small objects on public remote sensing image datasets,  verifying  its effectiveness and robustness under complex background conditions.

Key words: remote sensing image, rotated small object detection, dual-domain query enhancement, Transformer model