To overcome low efficiency and accuracy of existing forest fire detection algorithms, this paper proposes a network model to enhance the real-time and robustness of detection. This structure is based on the YOLOv5 target detection algorithm and combines the backbone network with The feature extraction module combines the attention module dsCBAM improved by depth-separable convolution, and replaces the loss function CIoU of the original model with a VariFocal loss function that is more suitable for the imbalanced characteristics of positive and negative samples in the forest fire data set. Experiments were conducted on a self-made and public forest fire data set. The accuracy and recall rate of the model can reach 87.1% and 81.6%, which are 7.40% and 3.20% higher than the original model, and the number of images processed per second reaches 64 frames, a growth rate of 8.47%. At the same time, this model was compared horizontally with other improved methods. The accuracy, recall rate and processing speed were all improved in the range of 3% to 10%. The effectiveness of the improved method in this article was verified, and the external perception level of the forest fire scene was deeper.