Object detection is one of the key tasks of UAVs. As the changes of UAV flight altitude and scene, the scale of detected objects is uneven and the proportion of small objects is high, which brings great challenges to the object detection under the UAV viewpoint. In order to address these problems, a reconstructed feature-based UAV image detection model RF-YOLOv8 was proposed. We change the traditional downsampling method in the spatial pyramid module to soft pooling comprehensively considering the global features to design the Spatial Pyramid Soft Pool-Fast and Efficient (SPSPFE) module, which improves the feature utilization and thus improves the detection effect of small objects. We have also introduced the Large Selective Kernel (LSK) module, which dynamically adjusts the receptive field in the feature extraction module to more effectively handle the differences in background information required for different objects. Additionally, we have designed the SCRELAN module by combining Spatial and Channel Reconstruction Convolution (SCConv) and RepNCSPELAN4, which reduces redundant calculations and promotes the learning of representative features. Furthermore, we have introduced a bounding box regression method, Shape-IoU, which focuses on the shape and scale of the bounding boxes, making the predicted boxes better fit the GT boxes. Experimental results show that RF-YOLOv8 achieves at least 5.3% improvement in mAP0.5 compared with other recent state-of-the-art methods, and achieves at average 10.2% reduction in parameters compared with the baseline YOLO model.