LLD-GAN: An end-to-end network for low-light image demosaicking

Demosaicking of low and ultra-low light images has wide applications in the fields of consumer electronics, security, and industrial machine vision. Denoising is a challenge in the demosaicking process. This study introduces a comprehensive end-to-end low-light demosaicking framework called LLD-GAN (Low Light Demosaicking Generative Adversarial Network), which greatly reduces the computational complexity. Our architecture employs a Wasserstein GAN framework enhanced by a gradient penalty mechanism. We have redesigned the generator based on the UNet++ network as well as its corresponding discriminator, which makes the model learning more efficient. In addition, we propose a new loss metric grounded in the principles of perceptual loss to obtain images with better visual quality. The contribution of Wasserstein GAN with gradient penalty and perceptual loss function was proved to be beneficial by our ablation experiments. For RGB images, we tested the proposed model under a wide range of low light levels, from 1/30 to 1/150 of normal light level, for 16-bit images with added noise. For actual low-light raw sensor images, the model was evaluated under three distinct lighting conditions: 1/100, 1/250, and 1/300 of normal exposure. The qualitative and quantitative comparison against advanced techniques demonstrates the validity and superiority of the LLD-GAN as a unified denoising-demosaicking tool.