Deep gradual flash fusion for low-light enhancement

Abstract

In this paper, we propose gradual flash fusion, a new imaging concept that enables acquisition of pseudo multi-exposure images in a passive manner. This means that our gradual flash capture does not require any user-side manipulation (taking multiple shots or varying camera settings). Continuous high-speed capture naturally contains different intensities of flash in a single shooting. The captured gradual flash images, containing different information of the same scene, are fused to generate higher-quality images, especially in a low light scenario. For gradual flash fusion, we use a Generative Adversarial Network (GAN) based approach, where the generator is a tailored convolutional Auto-Encoder for image fusion. For the training, we build a custom dataset comprising gradual flash images and corresponding ground truths. This enables supervised learning, unlike most conventional image fusion studies. Experimental results demonstrate that gradual flash fusion achieves artifact-free and noise-free results resembling ground truth, owing to supervised adversarial fusion.