Forest Fire Detection Algorithm Based on DetNet-FPN Feature Fusion Network

The occurrence of forest fire has caused a large area of forest damage, casualties and economic losses, and forest fire detection is the key to the timely warning of fire. The problem of small target loss still exists in the forest fire detection algorithm using FPN network. In order to solve the problem of poor definition of object edge and loss of small target flame semantic information caused by 32-fold downsampling in FPN multi-scale feature fusion network, a forest fire detection algorithm based on DetNet-FPN feature fusion network was proposed.  The backbone network of the algorithm adopts DetNet59, which is specially designed for target detection task. The network is improved on the basis of ResNet50, and the sixth stage is added. In order to maintain the resolution of high-level feature map, downsampling is abandoned in the fifth and sixth stages. Furthermore, dilated convolution is used to replace the original bottleneck structure with 3x3 convolution to enlarge the receptive field of feature map, thus improving the detection ability of small scale targets. Experimental results show that compared with FPN algorithm, the average accuracy of the proposed algorithm is improved by 2.70%, and the accuracy of small target is improved by 2.3%, which has good detection effect in various scenarios.