Fire Image Augmentation based on Diverse Alpha Compositing for Fire Detection

Fire image acquisition is dangerous and expensive. The lack of large-scale fire image dataset limits the application of data-driven methods on the fire detection task and the existing fire image augmentation methods are difficult to ensure the realism of generated flame images. To address this problem, this study presents a diverse alpha-compositing-based fire image augmentation method (DACBFIAM), which can generate a large-scale fire image dataset from a small-scale one. The fire image is generated by compositing a flame image into other nonfire images according to alpha composition model. The flame image is extracted by the automatic flame extraction step, and the nonfire image is obtained from public image datasets. In addition, DACBFIAM are also applied to increase the diversity of the compositing fire images. Experimental results show that the performance of data-driven fire detection methods with DACBFIAM outperforms that with other fire image augmentation methods and that without image augmentation.