Experimental Investigation of the Pix2pixHD Model for the Improvement of the Fairly Substantial Quality of Low-Light Images

In image synthesis, conditional generative adversarial networks (cGANs) have been investigated numerous times and are used for diverse image generation applications. In this work, the domain of low-light image enhancement that involves transforming low-light images into realistic, bright images employing a cGAN is explored. The pix2pixHD model, a cGAN, has been tested, examined, and improved qualitatively and quantitatively for low-light image enhancement. The employed quantitative metrics include both a no-reference Image Quality Assessment (NIQA) metric such as entropy and full-reference IQA (FIQA) metrics such as MSE, PSNR, SSIM, and LPIPS. Additionally, the pix2pixHD model with different combinations of loss functions is being investigated. Furthermore, the efficacy of histogram-equalized versions of input images in generating well-illuminated images is investigated by employing the lighter pix2pixHD model. The qualitative and quantitative simulation results are also compared with state-of-the-art deep learning models, and this analysis showed that pix2pixHD can challenge the most recent deep learning models for low-light image enhancement, namely MIRNet, LLFlow, and MIRNetv2, with a minimum inference time of approximately 0. 01s on GPU.