The rapid increasing use of unmanned aerial vehicle (UAV) has driven a growing demand for robust aerial target detection. However, achieving both high detection accuracy and real time performance, particularly for small targets in complex environments, remains a significant challenge. This paper presents FLSNet, an enhanced Real Time Detection Transformer (RT-DETR) with three key improvements to enhance efficiency and accuracy. FLSNet integrates Frequency Adaptive Dilution Convolution (FADC) into its backbone, enabling dynamic dilation rate adjustment based on input feature frequency. Adopting Learned Positional Encoding (LPE) optimizes the Attention-based Intra-scale Feature Interaction (AIFI) module’s positional encoding, aligning positional info with small targets’ spatial distribution to boost positioning accuracy. Moreover, the novel Small Object Enhance Pyramid (SOEP) strengthens small target features via multi-scale fusion, outperforming conventional methods merely adding a P2 layer. Evaluations on VisDrone2019 show FLSNet’s efficacy: recall 0.412, mAP50 0.398, mAP50–95 0.229, with 74.7 FPS. These surpass YOLO and state-of-the-art (SOTA) methods, with manageable parameters and complexity, meeting real-time needs. FLSNet balances accuracy and efficiency for UAV small target detection, offering a robust solution.