An Improved YOLOv8s Model for Small Object Detection in UAV Aerial Images

Currently, when detecting small targets in images captured from aerial drones, the mainstream detection algorithms face a series of issues, such as missed detections, false positives, and an imbalance between precision and speed. Based on the YOLOv8s model, we propose an improved algorithm for small target detection in drone-captured images. Firstly, we've enhanced the network's ability to detect small targets by adding a dedicated detection layer to the neck of the network. We've introduced an SPD (Space-to-depth) layer in the shallower layers of the network, which enables the model to effectively extract the feature information of small targets present in the images. Within the C2f module of the network, we've integrated the Triplet Attention mechanism to suppress the impact of complex backgrounds on detection and enhances the model's performance in small target detection tasks. Finally, we have added the CPCA (Channel Prior Convolutional Attention) module to the neck network, which further enhances the quality of composite features of targets and reduces sensitivity to background noise. Through the design of corresponding experiments, the final results show that our proposed detection model has improved the mAP0.5 by 6.7% on the public dataset VisDrone2019 compared to the YOLOv8s, demonstrating the advantage of our model in small target detection.