MFN-Net: multi-scale adaptive fusion and flexible label assignment for small object detection in remote sensing images

Object detection in remote sensing images is important yet challenging due to the small size of objects and interference from complex backgrounds. To address these challenges, we propose a novel detection framework named MFN-Net. First, a Multi-scale Feature Enhanced Adaptive Fusion architecture (MFEAF) is introduced, which effectively captures fine-grained features of objects by incorporating multi-scale downsampling and an adaptive weight fusion mechanism, thereby enhancing the representational capability of small object features. Second, we design a Similarity Distance-based Flexible Label Assignment (SFLA) strategy specifically tailored for small objects. This strategy employs aspect ratio filtering and a dynamic matching mechanism to assign sufficient high-quality positive samples to small objects, thereby improving detection performance. Through experiments on the SODA-A and AI-TOD datasets, the proposed method shows advantages in small object detection. The average precision (AP) is improved by 1.9% and 1.1%, respectively, compared to current state-of-the-art methods. Especially in the detection of very small and small objects, the method achieves a more significant performance improvement, demonstrating its superiority in small object detection tasks.