Small Target Detection Algorithm for UAV Aerial Photography Based on Improved YOLO V8n

UAV aerial photography has good applications in various aspects such as traffic management and disaster response. For UAV aerial photography need to detect the image, the detection object is mostly small targets, the target scale changes, the standard object is easy to be obscured and other problems, which leads to the existing small target detection algorithm has a high leakage rate, high false detection rate, inaccurate positioning and other problems. This paper proposes a small target detection algorithm for UAV aerial photography based on the improved YOLO V8n. A new hybrid module that combines the C2f structure in YOLO V8n with the EfficientAttention attention mechanism is designed to improve feature extraction and expression learning capabilities. SPD-Conv is introduced in the backbone network so that it replaces all the downsampling modules and makes it possible to perform downsampling without losing information. Replacing the head structure, a dynamic head design, dyHead, is introduced, aiming to further improve the accuracy of target detection. Replacing the original non-extremely large value suppression with Soft-NMS in the image post-processing stage to reduce the target miss-detection due to the dense and occlusion problem. The experimental results on Visdrone dataset show that the improved YOLO v8n algorithm achieves an average accuracy of 38.8%, compared with the YOLO v8n algorithm, the mAP@50, mAP@95 and Precision are improved by 8%, 9% and 27%, respectively, and the rate of misdetection and omission is low, which meets the requirement of detecting low-altitude flying object targets in complex background.