ResFusionYOLO: a distillation-based enhancing network by fusing object detection and super-resolution tasks for remote sensing imagery

Capturing high-frequency details of small objects in remote sensing images (RSI) is challenging. Existing solutions often employ complex neural networks to directly fit the task objectives, leading to high computational load. In this paper, we propose a fast yet accurate object detection (OD) method for RSI, named ResFusionYOLO. Our method aligns high-frequency information (HF) from a super-resolution (SR) model with OD data, and enhances the detection accuracy by capturing cross-modal features without extra computational overhead. We design a SuperDistil mechanism (SD) is designed to distill HF from the SR auxiliary task to the decoder of OD. To prevent additional computation, SD is discarded during the inference stage. Furthermore, we propose a cross-modality residual connection module to propagate the multi-scale low-frequency information (LF) of OD to the SR decoder, enabling the reuse of LF. Meanwhile, we design a Multi-Frequency Fusion Module to alignment and integration cross-modal features, enhancing feature fusion. Finally, we utilize UltraRefine Module to further optimize the SR objective. Extensive experiments on the VEDAI dataset demonstrate that the proposed ResFusionYOLO significantly outperforms other state-of-the-art methods and exhibits strong versatility.