The efficiency of unmanned aerial vehicles (UAVs) is widely used in industry for target surveillance. However, the accuracy of UAV target detection is limited by the complexity of the background and the high number of small targets. To address these issues, we propose a You Only Look Once (YOLO) detection network with contrast learning and similarity feature fusion (YOLO-CS). For the problem of complex background, we design a target occlusion contrast module; this module prevents the model from detecting background noise as a target by improving the differences between the background and the target. And then, a similarity fusion module is proposed to address the issue of small target detection; this module leverages similarity to selectively fuse multiscale features and effectively avoid small-scale features being overwritten by large-scale features, resulting in the miss detection of small targets. The experimental results on the VisDrone2019 dataset and the UAVDT dataset show that all the proposed modules effectively improve the detection performance of the model, and the proposed YOLO-CS model outperforms other popular methods.