Abstract: In order to quickly and effectively detect fire in different scenes and avoid missing the best time for fire fighting, an improved VGG16 image recognition fire detection method is designed based on deep transfer learning. Collect photos of fire and no fire in different scenarios, use offline data enhancement methods to increase the number of samples, improve VGG16, and use transfer learning methods to train fire recognition models. The experimental results show that the improved VGG16 model has a 98.7% accuracy in classification and recognition of pictures with and without fire, which is better than the Resnet50 model and the Densenet121 model. It is proved that the method has high accuracy in identifying the situation of flames after the fire, and can detect the fire quickly and accurately. 

Key words: fire protection, fire detection, image classification, VGG16, deep learning