Aiming at the problems that small target features in remote sensing images are easily lost, easily affected by background noise and difficult to locate, this paper improves the YOLOXS target detection model. Firstly, the CBAM is improved by using the twodimensional discrete cosine transform and added to the backbone network to improve the attention of the network to small targets; secondly, a weighted multireceptive spatial pyramid pooling module is proposed to improve the perception ability of the model to multiscale targets, especially to smallscale targets. Thirdly, using the idea of crosslayer feature fusion, a crosslayer attention fusion module is proposed to retain as many small target features as possible in the deep structure; finally, EIoU loss is used to enhance the localization ability of small targets. As shown by extensive experimental analysis, the APs value of the improved model improves by 51% relative to the baseline model on the RSOD dataset and by 24% on the DIOR dataset, and the number of parameters increases by only 1.01 M. The detection speed reaches 93.6 fps, which meets the detection requirements of real-time. In addition, the improved model in this paper also has certain advantages over the current stateoftheart target detection models.