FGL-GAN: Global-Local Mask Generative Adversarial Network for Flame Image Composition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Collecting flame image datasets poses risks.
• Existing methods for compositing flame images often yield low-quality results.

Purpose of the Study:

• To propose a novel Generative Adversarial Network (GAN) for high-quality flame image compositing.
• To enhance the realism and utility of synthetic flame datasets for fire detection.

Main Methods:

• A Global-Local mask Generative Adversarial Network (FGL-GAN) with a hierarchical generator.
• Incorporation of fire masks and a novel data augmentation technique during training.
• Integration of contrastive learning to improve fitting speed and reduce blurriness.

Main Results:

• FGL-GAN significantly outperforms mainstream GANs in qualitative and quantitative evaluations.
• Ablation studies confirm the effectiveness of the hierarchical generator, fire mask, data augmentation, and MONCE loss.
• The method successfully generates high-quality flame halo and reflection with consistent global style.

Conclusions:

• FGL-GAN provides a robust solution for generating realistic synthetic flame images.
• The generated datasets can extensively support the training and testing of deep learning-based fire detection equipment.
• This approach mitigates the dangers and limitations of collecting real-world flame image data.