Abstract: In order to solve the problem of insufficient sample library of fire images in large-span spatial places, which leads to low accuracy of fire detection, a method of enhancing synthesized samples of large-span spatial fire is proposed. This method is based on an improved CycleGAN network, which constructs a flame block generation network and a fire image generation network. The flame block generation network generates flame block images, and the fire image generation network integrates flame blocks into designated areas of the engineering scene to generate fire samples that can be used for object detection models. The results show that the synthesized samples generated by the model improved the performance of the object detection model significantly, with a increase of 25.11% in mAP@0.5, verifying the effectiveness of the data augmentation method. Therefore, by generating high-quality and diverse large-span fire images, the problem of insufficient fire sample libraries has been effectively alleviated, providing strong support for the development of artificial intelligence based image fire detection algorithms for specific scenarios.

Key words: generative adversarial network, dataset augmentation, fire detection, sample synthesis