In order to reduce the false alarm phenomenon of smoke detection caused by cloud and fog, a smoke detection method based on fusing multiple network models is proposed. On the basis of using VGG16 network to extract the detailed features of smoke, it is fused with the ResNet50 network feature extraction layer to extract more subtle features. The skip connection mechanism is used to transfer the image information to the deeper layer of the neural network, in order to avoid the loss of important features of smoke image and solve the under-fitting problem caused by the gradient disappearance.The training process adopts the feature transfer learning method based on isomorphic space to solve the small sample training problem, retrain in the new target detection field, better integrate the network model, rebuild the output detection structure of the whole connection layer, and adopt the random inactivation method to improve the generalization ability of the model.Experimental results show that, compared with the current popular deep convolutional network, this method has lower false alarm rate and higher accuracy and recall rate.