With the gradual maturity of the target detection algorithm based on deep learning, its deployment on UAV has become a hot topic.Regarding the problems of low detection accuracy caused by many small targets and easy occlusion in UAV aerial images, complex detection scenes and large scale variability, an improved algorithm based on the proposed S-YOLOv4 is introduced.Firstly, SENet attention mechanism is added on the original feature extraction network structure to improve the models ability to concentrate on useful information and enhance inter-channel attention.Secondly, a new detection layer with a resolution of 160×160 is added to refine the grid for better detection of small targets.Finally, the loss function is improved, and the class smoothing label is applied to the classification loss to reduce the penalty of negative samples and improve the generalization ability of the model.Compared with that of the original algorithm, the mAP of the proposed algorithm is improved by 3.4% under the real-time detection speed.