RS-Detr: Detail Feature Extraction and Reassembly Modules for Fine-Grained RSIs Object Detection

Object detection in remote sensing images(RSIs) faces challenges due to small target sizes and multi-scale targets within the same range. Current RSI detection architectures use strided convolutions or pooling layers for higher-level features, causing a loss of detail information and performance degradation. Furthermore, low resolution of RSIs equailly complicates semantic information extraction. To overcome the issue of detail information loss during high-level feature extraction and the challenge of semantic mining in low-resolution images, we propose RS-DETR, which incorporates the Detail Feature Extraction Module (DFEM) and the Deep Feature Reassembly Module (DFRM). DFEM improves feature extraction accuracy and effi-ciency through positional encoding, feature map reconstruction, and discrete wavelet transform, capturing fine-grained and multi-scale information while integrating global information with deep convolution and cross-attention. DFRM, a lightweight upsam-pling module, uses depthwise separable convolutions to generate semantic masks for adaptive upsampling, achieving traditional upsampling effects with minimal computation. We also employ the EIoU loss function for faster model convergence. Experiments on the FAIR 1M and DIOR-R datasets validate our method, demonstrating its effectiveness in handling low-resolution remote sensing data and small object recognition.