A fire monitoring and alarm system based on channel-wise pruned YOLOv3

Fire detection and alarm system is fully concerned for safety. And convolutional neural network (CNN) has been introduced into fire/smoke detection based on video/image understanding. However, the samples of the existed public fire/smoke data sets are not enough to train very deep CNN. And the generalization abilities of the existed methods are limited. Therefore, a fire monitoring and alarm system (FMAS) based on channel-wise pruned YOLOv3 is proposed, and a big fire and smoke image data set has been collected in this paper. YOLOv3 with Darknet-53 is deep and its generalization ability has been demonstrated in general objects detection. But it has massive parameters, which may hinder its applications in fire monitoring systems with restricted computation resources. Thus, channel-wise pruning technology is introduced to reduce the number of parameters while bringing a slight drop of accuracy. Moreover, OHEM technology is proposed to improve the detection accuracy further. Multiple comparison experiments on the homemade data set and the public data sets have demonstrated that the proposed channel-wise pruned YOLOv3 with OHEM can achieve satisfactory accuracy with low calculation capacity after squeezing parameters.