Dual-Domain Feature-Guided Task Alignment for Enhanced Small Object Detection

Small object detection is a critical challenge in Unmanned Aerial Vehicles (UAVs) due to the limited pixel representation of small objects and the impact of successive pooling operations, which frequently results in the disappearance of small objects within intricate backgrounds. To tackle this issue, we propose the Small Object Enhancement Pyramid (SOEP) module, which first transforms feature representations (i.e., in the spatial domain) into the frequency domain to better capture small objects typically characterized by high-frequency components. These feature representations are then fused in the spatial domain using a frequency-based attention map, enhancing small object representations by integrating information from both complementary domains. Furthermore, we introduce a Task Aligned Head (TAH) that integrates classification and localization tasks interactively, reducing the misalignment that occurs when these tasks are learned independently, particularly in the context of small objects. Experimental results on the Visdrone dataset verify that our proposed method (D²FTA) outperforms the baseline method by 12.7%, 14.19% on mAP0.5 and mAP0.5:0.95.