L2BGAN: An image enhancement model for image quality improvement and image analysis tasks without paired supervision

Jhilik Bhattacharya, Gianni Ramponi, Leonardo Gregorat, Shatrughan Modi

Abstract:

The paper presents an image enhancement model,
L2BGAN, to translate low light images to bright images
without a paired supervision. We introduce the use of geo-
metric and lighting consistency along with a contextual loss
criterion. These when combined with multiscale color, tex-
ture and edge discriminators prove to provide competitive
results. We perform extensive experiments on benchmark
datasets to compare our results visually as well as objec-
tively. We observe the performance of L2BGAN on real time
driving datasets which are subject to motion blur, noise and
other artifacts. We further demonstrate the application of
image understanding tasks on our enhanced images using
DarkFace and ExDark datasets.

One-sentence Summary: Image Enhancement without paired supervision