Generating High-Resolution Fire Images with Controllable Attributes via Generative Adversarial Networks

Obtaining realistic fire images using deep learning models and recent versions of Generative adversarial networks (GAN) has been proven to be a difficult task due to the unnatural appearance of the generated results. This paper provides a novel approach based on StarGANv2 to generate fire kernels from any input provided as a reference. In addition, a deep learning-based image blending technique performs the migration of the fire kernels to the target scenes. By using any input as a reference, the generated fire image could be controlled to accommodate different environmental factors, resulting in a diverse but equally pseudo-real synthetic dataset. The proposed method generates images that achieve better FID and LPIPS values than StarGANv2 for both a public dataset (AI Hub) and a privately-owned dataset (Visionin). In addition, YOLOv4 is used as a fire detection model to evaluate the synthetic data on improving the performance of the detected network. Compared to the model trained on the real data, the model trained on the combined dataset outperforms 2%similar to 14% higher.