Flash image enhancement via ratio-log image translation to ambient images

Abstract

To illuminate low-light scenarios in photography, photographers usually use the camera flash, this produces flash images. Nevertheless, this external light may produce non-uniform illumination and unnatural color of objects, especially in low-light conditions. On the other hand, in an ambient image, an image captured with the available light in the ambient, the illumination is evenly distributed. We therefore consider ambient images as the enhanced version of flash images. Thus, with a fully convolutional network, and a flash image as input, we first estimate the ratio-log image. Then, our model produces the ambient image by using the estimated ratio-log image and ash image. Hence, high-quality information is recovered with the flash image. Our model generates suitable natural and uniform illumination on the FAID dataset with SSIM = 0:662, and PSNR = 15:77, and achieves better performance than state-of-the-art methods. We also analyze the components of our model and how they affect the overall performance. Finally, we introduce a metric to measure the similarity of naturalness of illumination between target and predicted images.