Forest Fire Detection Based on Improved YOLOv5

Traditional forest fire detection based on sensing devices has a small range, low sensitivity and high false alarm rate, while existing forest fire detection methods based on deep learning have poor identification effect and inaccurate identification of small smoke and fire targets in practical application. To solve the above problems, this paper proposes an improved YOLOv5 network forest fire detection method. Firstly, the image is divided into multiple grids. The adaptive anchor setting optimization method is used to generate the adaptive anchor parameters through the training of neural network, so that the network has certain pertinence in forest fire detection. Then CSPDarknet-53 was used to extract the feature network and train to get the optimal weight model, so as to detect the fireworks target in the image. In order to solve the problem of actual forest scenarios, the accuracy of the model is improved by building a high-quality forest fire data set. Using the CIoU loss function to improve the network detection capability, and the combination of Soft-NMS and DIoU-NMS post-processing to improve the inhibition effect of redundant bounding box and reduce the recall rate. The experimental results show that the improved algorithm has a higher detection speed than the traditional network in forest fire detection, and the model has a good multi-scale fire detection performance, which has a certain application value in the real scene forest fire detection.