MDADet: A Multimodal Dynamic Adaptation Framework for Efficient Small Object Detection in Aerial Images

In this paper, we propose a multimodal dynamic adaptive detection framework tailored for small object detection named MDADet. Concretely, we utilize a Dynamic IoU-Centric Slicing-based Data Augmentation (DICSA) strategy to prioritize high-IoU regions during training. The strategy effectively eliminates redundant background information and significantly accelerates model convergence. Additionally, the Robustly Optimized BERT Pretraining Approach (RoBERTa) encodes bounding box annotations into semantic embedding, which are fused with image features via a transformer to generate multimodal representations for small object recognition. The knowledge distillation is utilized to transfer capabilities from the multimodal teacher model to a lightweight multimodal student model, reducing parameter scale and improving inference speed. During fine-tuning of the single-modal student model, the transformer encoder is frozen, and a lightweight feature pyramid integrated with Pixel-Shuffle and hierarchical detection heads is incorporated, ensuring robust performance even without textual input. Experimental results compared with other methods demonstrate the effectiveness and advancement of MDADet, achieving 81.07% mAP on DOTA 1.0, 86.76% on VEDAI, 73.55% on DIOR and 97.61% classification accuracy on NWPU VHR-10, with a model size of only 37.8M parameters.