An Improved Multi-Scale Fire Detection Method based on Convolutional Neural Network

Fire is a major disaster in the world, and the fire detection system should accurately detect the fire in the shortest time to reduce economic loss and ecological damage. Traditional sensors are still widely used in a large number of applications, but they do not perform well in remote high-dome environments or the early stages of low-flame fires, and now the method of using image and video to predict fire is becoming more and more popular. This paper proposed an improved YOLOv4 fire detection method based on Convolutional Neural Networks (CNN). We improve the accuracy of the model through the self-built high-quality fire dataset, use the changed loss function to improve the detection ability of small-scale flames, and combine the Soft-NMS post-processing and DIoU-NMS post-processing to improve the suppression effect of the redundant Bounding box and reduce low recall rate. The experimental results of the model on our dataset show that the model has an excellent performance in fire detection and can detect multi-scale fire in real-time.