To solve the problem that the detailed features of the high-level feature map are weakened and the low-level features of the smoke image are lost, an improved YOLOv4 algorithm that combines the detailed features and the attention mechanism is proposed. The detail feature fusion module is designed, and the low-level features in backbone network are introduced into high-level features to obtain the fusion feature map with extensive multi-scale information. Then, a hybrid attention mechanism in two separate dimensions of channel and spatial is adopted, to reassign the weight of the fusion feature map. The smoke target features are enhanced the background features are suppersed, hence the proposed algorithm is robust in feature expression. The experimental results show that the average precision, precision and recall rate of the algorithm in this paper are increased by 4.31%, 1.21% and 9.86% respectively compared with the YOLOv4 algorithm, while maintaining a high detection speed. The proposed algorithm can effectively extract the overall features of smoke targets, and is suitable for smoke detection tasks in complex backgrounds.