Unmanned aerial vehicle(UAV) images have distinctive characteristics such as significant scale variance, background interference, and fuzzy objects, which brings many challenges for object detection algorithms. To address these issues, an efficient algorithm for small object detection in UAV images is proposed. Firstly, parallel atrous convolution(PAC) is designed to increase the receptive field and maintain detail resolution by utilizing atrous convolution. Then, an attentional upsampling and downsampling module is designed to selectively enhance the extracted features by using gate mechanisms. Finally, in combination with tiny object prediction head (TOPH), a parallel atrous convolution attention pyramid network(PACAPN) is designed to fuse multi-scale features. On the VisDrone2023 dataset and the DOTA dataset, the proposed algorithm achieves higher mean average precision(mAP) in both small object detection in comparison other algorithms. Compared with the baseline method, the mAP is increased by 7.3 %, the number of parameters is reduced by 0.58 M, and the FPS(frame per second) is increased by 11.2 to 43.5, which verifies the effectiveness of the algorithm. On the ExDark dataset, the proposed algorithm achieves higher mAP in comparison low-light enhancement models and dark detectors. Compared with the PE-YOLO, the mAP is increased by 2.4 %, which verifies the robustness and practicality of the proposed algorithm.