Objective This study aims to optimize and improve the benchmark YOLOv5s algorithm to solve the problems of low accuracy for small targets and complex models in complex scenarios during unmanned aerial vehicle UAV remote sensing image detection. Methods The lightweight convolutional block attention module CBAM and a stacking fusion strategy were introduced to reconstruct the backbone C3 template. This approach enhanced feature extraction ability of the network while reducing the number of parameters. In the feature fusion layer cross-level feature fusion was introduced to avoid feature loss and an upsampling operator was incorporated during the fusion process to reduce the complexity of the model and fuse more feature information. Additionally the integrated loss function ILF was adopted as the bounding box loss function to enhance localization accuracy for targets. Results In the VisDrone2019 dataset test the improved algorithm achieved a 5. 8% increase in mean average precision mAP compared to the original algorithm. The model size was reduced to 6. 2 MB a significant reduction from the original model while the number of parameters decreased by 44. 7%. Moreover the proposed method achieved higher detection accuracy compared to mainstream detection approaches. Conclusion The proposed improved method significantly reduces the complexity and computational cost of the model while maintaining the detection accuracy and achieves fast and accurate recognition of small targets in remote sensing images. For small target detection tasks this method has good performance and is of great significance for realtime and resource-constrained application scenarios.