Small target detection in drone aerial images based on feature fusion

The use of object detection technology in unmanned aerial vehicles is a crucial area of research in computer vision. Aerial images captured by drones exhibit differences in object shape and size compared to traditional images, which can cause object detection algorithms to miss or misidentify small targets. This paper makes improvements based on the YOLOv5 algorithm. The algorithm introduces a small target detection layer to improve the model's detection capability at different scales. Cross-channel fusion module and multi-level feature fusion downsampling module are added to obtain more comprehensive context information. This makes the network pay more attention to the important features of small targets. Additionally, the classification task and regression task of the detection head are decoupled to speed up the model's convergence and improve detection accuracy. Finally, a new loss function is proposed to further improve the accuracy and convergence rate of the detector. The algorithm is evaluated on the VisDrone2019 dataset and compared with the YOLOv5s algorithm. The results show an improvement of 4.7% in mAP0.5, 3.0% in mAP0.5:0.95, 3.6% in precision, and 6.4% in recall. At the same time, the algorithm was evaluated on the DIOR dataset, and mAP0.5:0.95 improved by 1.5%.These findings demonstrate the algorithm's effectiveness in detecting small targets in aerial images captured by drones.