DLA-Deformable DETR: Dynamic Layer Assignment Deformable DETR for Remote Sensing Images

The rapid advancement of remote sensing technology has made extracting accurate information from remote sensing images increasingly challenging due to multi-scale object distribution, object sparsity, and complex backgrounds. To address these challenges, we propose the Bi-directional Cross Stage Partial Feature Pyramid Network (BCSPFPN), which enhances the integration of high-level semantic features and low-level fine-grained features. We also introduce the Dilated Convolution Pyramid Module (DCPM) to improve the model's ability to capture high-level semantic information and detect small objects. Additionally, the Dynamically Allocated Deformable Decoder (DA-Deformable Decoder) dynamically allocates features from different levels to the decoder, addressing early-stage feature extraction issues. Experiments on the NWPU VHR-10 and RSOD datasets demonstrate that our method achieves state-of-the-art performance in object detection.