Fire Alarm for Video Surveillance Based on Convolutional Neural Network and SRU

At present, all CNN-based fire identifications identify whether a fire is blazed with a single frame image, all of which have low accuracy under a strong interference or complex backgrounds such as flickering light or a high-brightness background. To address the issue, this paper proposes a neural network model which combines CNN with SRU. The scene content is extracted through CNN and combined with the dynamic characteristics of the flame extracted by SRU to improve the accuracy of fire alarm. There are three models proposed in the paper, including Resnet18+SRU, Resnet34+SRU and resnet18(Maxpool)+SRU. The models were validated on a test set containing intensive indoor environmental interferences and compared with CNN-based single-frame and multi-frame fire identification methods. The results show that the proposed methods are more accurate than the single-frame CNN fire recognition method by more than 25% and are more suitable for indoor environment fire alarms.