Smoke Recognition Based on Deep Transfer Learning and Lightweight Network

Fire is one of the most common catastrophic events, and smoke is one of the most significant characteristics of fire. Accurate and real-time smoke detection is an important means to reduce fire risk. In this paper, we construct AlexNet, VGG16, GoogLeNet Inception V3, ResNet50 and lightweight MobileNet network models to identify smoke, and then add transfer learning methods to make a contrast. We find that pre-training VGG16 network which has a big module and many parameters is most suitable for this identification and the accuracy of it is about 0.01 higher than that of others. Pre-training MobileNet network model whose size and number of model parameters are only about 1/15 of AlexNet network can achieve similar accuracy to AlexNet. This research is of great help to improve the training speed and reduce the cost of updating the model and it provides the possibility for the deployment of the model on the devices with limited computing power.