Dynamic Fusion Label Assignment Network for Remote Sensing Object Detection

Object detection in remote sensing images faces significant challenges posed by tiny objects, which are often overwhelmed by background noise. These tiny objects exhibit significant scale differences compared to larger objects, making it difficult to simultaneously consider both large and small objects during label assignment. In this paper, a novel dynamic fusion label assignment network (DFLAN) is proposed to address these problems. Firstly, to effectively extract features of tiny objects in background noise, we introduce a novel feature selection interactive pyramid network. Secondly, a novel dynamic fusion label assignment algorithm is developed, which achieves a collaborative approach to label assignment for both tiny and large objects. Finally, a new decoupling detection head is proposed to prevent task coupling from interfering with the already weak features of tiny objects. The proposed DFLAN method achieves state-of-the-art performance on two widely-used datasets: DOTA-v1.0 (79.42% mAP), HRSC2016 (98.86% mAP) and DIOR-R (67.68% mAP).