Small object detection in Unmanned Aerial Vehicles’ (UAVs) aerial images faces challenges such as low detection accuracy and complex backgrounds. Meanwhile, it is difficult to deploy the object detection models with large parameters on resource-constrained UAVs. Therefore, a lightweight small object detection model EMFE-YOLO is proposed based on efficient multi-scale feature enhancement by improving YOLOv8s. Firstly, the Enhanced Attention to Large-scale Features (EALF) structure is applied in EMFE-YOLO to focus on large-scale features, improve the detection ability to small objects, and decrease the parameters. Secondly, the efficient multi-scale feature enhancement (EMFE) module is integrated into the backbone of EALF for feature extraction and enhancement. The EMFE module reduces the computational cost, obtains richer contextual information, and mitigates the interference from complex backgrounds. Finally, DySample is employed in the neck of EALF to optimize the upsampling process of features. The EMFE-YOLO is validated on the VisDrone2019-val dataset. Experimental results show that it improves mAP50 and mAP50:95 by 8.5% and 6.3%, respectively, and reduces the parameters by 73% compared to YOLOv8s. These results demonstrate that EMFE-YOLO achieves a good balance between accuracy and efficiency, making it suitable for deployment on UAVs with limited computational resources.
Yang et al. (Thu,) studied this question.