A New Color and Texture Augmentation Method of Flame Images Based on Diffusion Models

In order to simulate the unique color and texture changes of fuel flames under different lighting environments and achieve data augmentation, a flame coloring framework that expands its diversity by adding diverse colors and textures to flame images is designed in this paper. By using a diffusion model to map noise to the texture domain and using a textureless flame color image as the control condition, a one to many mapping from the image domain to the flame image domain is successfully achieved. This method can not only adjust the color and texture intensity of flames based on lighting intensity, but also generate flame images with variable textures under the same lighting conditions. The model proposed in this article has important reference value for improving the performance of video fire detection methods based on deep learning.