Abstract: In order to improve the accuracy of fire detection in actual scenes, a fire detection algorithm was proposed based on Anchor-free structure according to the characteristics of fire. Firstly, the ResNet block of feature extraction network was designed as a multi-branch structure, and the attention block designed according to the flame features was embedded on the basis to extract the more expressive features. And then, a subpixel fusion was added to enhance the expression ability of multi-scale features by utilizing the abundant feature information in high-level channel. Finally, a feature enhancement module was designed to enhance the top-level feature representation and make better use of global spatial information. Adaptive label assignment was introduced to enhance the learning ability of the extraction network to flame features. Retreating carefully to the boundary condition with improved GIoU Loss function, the detection accuracy of this algorithm can reach up 94.9% on the self-built data set, also show a good detection effect on the public data set. The experimental results show that the algorithm model can provide a high detection accuracy and strong anti-interference ability. In addition, this algorithm can also provide a better detection effect on multi-scale flame under complex background, suitable for fire detection in various environments, can meet the needs of actual fire detection tasks.