The traditional network calculation complexity of the safety helmet detection algorithm at the power construction site is high,and there are problems such as missing detection for distant targets and dense groups in complex scenarios.This paper proposes an improved lightweight YOLOv5s-GCAE algorithm,in which the backbone network first uses the deep separable convolutional GhostConv in the GhostNet network to reduce the amount of computation and parameters of the network.Secondly,the CA attention mechanism is embedded in the feature extraction stage,which fills the lack of accuracy when introducing lightweight networks.The ASFF network is introduced to effectively fuse multi-scale features,improve the rich semantic feature representation of the model,and make the network better adapt to complex power construction sites.Finally,the loss function EIOU is introduced to promote the network to focus on high-quality anchor points to improve the accuracy of helmet detection in complex scenarios.In this paper,a safety helmet wearing detection dataset containing 9326 open-source images and self-collected images is constructed.Experimental results show that the safety helmet detection accuracy of the algorithm is 93.4%, which is 2.1%higher than that of the YOLOv5s algorithm,which meets the accuracy requirements of safety helmet detection in power scenarios.