GCGE-YOLO: Improved YOLOv5s Algorithm for Object Detection in UAV Images
Gang Xiong, Juntong Qi, Mingming Wang, Chong Wu, Hailiang Sun
Abstract
Current general-purpose object detection algorithms cannot fully extract the features of small objects in UAV aerial images, and they usually suffer from problems such as large size, complex structure and large computation, making it difficult to deploy real-time object detection algorithms with high precision in UAV platforms. In response to the above problems, this paper intends to optimize the YOLOv5s algorithm and proposes a GCGE-YOLO algorithm that is lightweight and offers improvements in precision over the baseline. Firstly, the GhostNet is introduced to compress the Backbone of YOLOv5s to reduce the complexity and computation of the network model; secondly, the coordinate attention (CA) is introduced in the Backbone to make the model pay more attention to the main information to improve the precision; thirdly, GSConv combined with GhostNet is used as the Neck to fuse the detail information in the shallow feature map and enhance the feature extraction ability for small objects; finally, EIoU is used as the loss function of the algorithm to improve the object border localization precision while enhancing the speed of bounding box regression. In this paper, experimental validation is conducted on the public dataset VisDrone2021 and the algorithms are deployed on the embedded platform NVIDIA Jetson Xavier NX. The experimental results show that compared with the YOLOv5s algorithm, the GCGE-YOLO algorithm improves the FPS by 41.4%, compresses the model parameters by 39.4%, reduces the model size by 37.5%, reduces the computation by 31.7%, and improves the mAP by 1.5%. In all, the GCGE-YOLO algorithm achieves good improvements in all aspects and can meet the requirements of real-time object detection for UAV platforms.