YOLOv5-CSF: an Improved Deep Convolutional Neural Network for Flame Detection

Fire is a multiple and destructive disaster that usually results in great loss of life and property. Therefore, early detection of fire can help minimize mortality and reduce the risk to ecosystems and property. We propose a novel high-precision flame monitor, YOLOv5-CSF, that can monitor fires using existing monitoring equipment. Based on the YOLOv5 detector, we introduce a coordinate attention mechanism from the backbone network to obtain the relationship between flame position information and channel information in an efficient way which increase the feature expression of the backbone network. The swin transformer block is introduced in the neck network to expand the perceptual field of the network model and improve the flame feature extraction capability. The adaptive spatial feature fusion module is introduced in the head network to strengthen the multi-scale feature fusion of flame features and reduce false alarms. Compared with the original YOLOv5l, the average accuracy of the model is improved by 4.1%. Compared with seven other advanced flame monitors, the proposed algorithm has the highest average accuracy in the flame dataset and the proposed method is quite effective.