Forest fire detection algorithm based on Improved YOLOv5

Abstract Aiming at the problem that the existing target detection algorithms are difficult to detect sudden fire with high precision in real-time in the complex forest environment, an improved forest fire detection algorithm based on YOLOv5 is proposed by us. Firstly, the k-means clustering algorithm based on 1-IoU distance is used to preprocess the forest fire data set. Secondly, the Transformer attention mechanism is added to the model backbone network to improve the detection accuracy, and the regression box loss function GIoU-NMS is changed to EIoU-NMS. Finally, the network structure is changed, and the deep separable convolution is added to reduce the parameters. Comparing the algorithm we propose with the original YOLOv5 algorithm and the mainstream target algorithm, the experimental results show that the algorithm can have more efficient and high-precision detection results in fewer samples and complex environments and prove its feasibility and effectiveness.