LIGHTWEIGHT FOREST FIRE DETECTION BASED ON DEEP LEARNING

Forest fire detection is a challenging problem in computer vision. In this paper, we build a challenging fire dataset which contains images of fire, smoke, and red leaf to better simulate the real forest environment. We propose a lightweight network structure, YOLOv4-Light, for forest fire detection. The original YOLOv4's backbone feature extraction network is replaced by MobileNet, and PANet's standard convolution is replaced by depthwise separable convolution, which improves the detection speed and makes it more suitable for embedded devices. We also adjusted the YoloHead according to the relationship between smoke and flame to reduce the missing rate and false rate. The experimental results show that our YOLOv4-Light achieves good performance for forest fire detection, at the same time, our YOLOv4-Light achieves higher FPS and the model size is reduced by 4 times compared with other algorithms, which makes it easier to implement on embedded devices.