To improve the precision and recall rate of fire smoke detection model in multi-scene fire smoke detection applications, and avoid the tedious manual smoke feature extraction process, a fire smoke detection model is proposed which is based on convolutional neural network YOLOv4. In the last layer of the backbone network, four different scales of maximum pooling are added: 13×13, 9×9, 5×5, and 1×1. The multi-scale feature fusion uses PANet (Path Aggregation Network) to improve network feature extraction capabilities. In addition, a channel attention network is added to the network prediction head to enhance the ability of the YOLO Head to extract effective smoke information. For the fire smoke data set, the size of the candidate frame is clustered using the K-means algorithm to get a size closer to the fire smoke data set. Due to the identification of smoke, the loss function is simplified, the classification error is eliminated, and the algorithm converges faster. Data enhancement methods such as image flipping and random erasure are used in the training phase to reduce the risk of overfitting. Experimental results show that the fire smoke detection model has excellent performance. Its precision can reach 92.5%, Recall can reach 87.7%, and the detection speed can reach 5.1 frames/s, which improves the performance of fire smoke detection model in multi-scene fire smoke detection applications.