A Retinex based GAN Pipeline to Utilize Paired and Unpaired Datasets
for Enhancing Low Light Images
Harshana Weligampola 1* , Gihan Jayatilaka 1 , Suren Srithraran 1 , Roshan Godaliyadda 2 , Parakrama Ekanayaka 2 ,
Roshan Ragel 1 and Vijitha Herath 2
1
Department of Computer Engineering, Faculty of Engineering, University of Peradeniya, Sri Lanka
2
Department of Electrical and Electronic Engineering, Faculty of Engineering,
University of Peradeniya, Sri Lanka
*E-mail: [email protected]
Abstract: Low light image enhancement is an important challenge for the development of robust computer
vision algorithms. Classical approaches for this problem have been unsupervised whereas the deep learning
approaches have mostly been based on supervised learning using either paired or unpaired dataset. This paper
presents a novel deep learning pipeline that can learn from both paired datasets and unpaired datasets. The
proposed model employs CNN and GAN to optimize and minimize the standard loss and the adversarial loss
respectively. Cycle consistency loss and a patched discriminator are utilized to further improve the performance.
Finally, the paper presents an ablation study on the functionality and the performance of different components
and hidden layers in addition to the analysis of the full pipeline by a broad collection of visual examples.
Key Words: low-light image enhancement, retinex theory, generative adversarial networks, cycle consistency
11