Aiming at the problems of complex smoke occurrence scene and low accuracy in smoke, an improved YOLOv5s smoke and flame detection method is proposed. Firstly, in order to solve the problem that the Neck feature fusion of smoke and flame is not accurate and the effect is poor, a new channel attention mechanism, Scoring module, is proposed to score features of each channel. Features with high scores are selected for feature fusion and features with low scores are filtered to avoid introducing too many redundant features. On the premise of not increasing too much computation burden, the module can enhance feature fusion ability and detection accuracy. Then, in order to improve the prediction ability of the Head layer, α-EIOU is used to replace GIOU as the prediction box regression loss to improve the prediction accuracy of the prediction box. Finally, the improved Mosaic data enhancement method is used to solve the problem of small data set and single data form, expand the sample data, and improve the generalization ability of the model. As a result, the mean average precision of the modified YOLOv5s model is improved by 4.7%, while the detection speed reaches 212 frames per second. Meanwhile, it performs well in the comparison experiment with other improved YOLOv5s. It achieves good detection effect in the image with complex environment, and can meet the task of smoke and flame detection in complex environment.