A More Lightweight and Accurate Object Detection Network for UAV Aerial Images

Object detection technology plays an important role in the application of unmanned aerial vehicle (UAV). There are many small targets, multiscale targets, and noise interference of complex environments in UAV aerial images, which often lead to false and missing detection. To solve the above problems and reduce the computational burden in practical application, this paper proposes a more lightweight and accurate object detection network based on YOLOv8 for UAV aerial images. We redesign an efficient feature extraction network based on MobileViT, which can simultaneously extract global and local information in complex backgrounds, and learn better feature representations with less computational cost. Besides, a feature fusion named DSCFF that can fully interact the semantic information with the spatial information is proposed, enhancing the detection ability of various scales, especially small targets. Finally, we perform the lightweight design of the decoupled head to reduce the increased computation due to the decoupled structure. Experiments on the VisDrone2019-DET dataset indicate the proposed method raises mAP50 by 2.2% and 2.4% on the validation set and test-dev set respectively, while its parameters and computation are reduced by 6.1 M and 4.3 G respectively. The experimental results on the DOTA-v1.0 dataset further prove the effectiveness of the proposed method.