An aerial image detection algorithm named FDL-YOLO based on multi-scale feature aggregation was proposed based on YOLOv11 aiming at the problem of high miss detection rate caused by feature confusion of dense small objects in unmanned aerial vehicle (UAV) aerial image detection task. A feature extraction module C3k2-FDCN based on frequency dynamic convolution was designed. The sensitivity to high-frequency detail features of small objects was enhanced by dynamically adjusting the frequency response characteristic of convolution kernel. A dynamic adaptive scale integration module was adopted to reconstruct the feature pyramid structure and add a new 160×160 high-resolution feature map. Then adaptive deep fusion of cross-scale features was realized. A LoG-Stem layer was introduced at the network input to enhance edge features of input image and strengthen target contour information. A shallow detail fusion module was constructed at the front end of the detection head to alleviate feature confusion in dense object scenario through dual channel-spatial attention mechanism. A dependency graph-based structured pruning strategy was employed to optimize model parameters while retaining key feature channel. Then a balance between detection accuracy and lightweight deployment was achieved. The experimental results on the VisDrone2019 dataset showed that the improved algorithm increased accuracy by 5.9%, recall by 8.2%, and mAP50 by 8.6% to 47.2%, with 70% reduction in parameters. The algorithm can effectively cope with the challenges in UAV aerial image object detection task.