Integrating Graph Convolution Into a Deep Multilayer Framework for Low-Light Image Enhancement

Digital camera sensors may struggle to capture images in low-light environments, resulting in lower brightness and contrast levels, color degradation, undesirable characteristics, and noise. Such reduced-quality images adversely affect the performance of computer vision algorithms. With the advancement in deep neural networks, many methods have emerged for improving images taken in dim lighting conditions. While the current practices prioritize enhancement, they often come at the cost of increased noise levels, making them impractical for real-world deployment. In this letter, a deep-learning architecture for the enhancement of low-light images is proposed. The suggested model is optimized through a deep multilayer framework. The search for similar nonlocal neighborhoods is modeled with the graph convolutional network. It uses only 0.32 M trainable parameters and has an average inference time of 22 ms. The experiments encompass three dataset combinations and incorporate quantitative comparison metrics, such as peak signal to noise ratio (PSNR), MSE, structure similarity index measure (SSIM), and multi-scale ssim (MSSSIM). In total, nine distinct approaches to low-light image enhancement were investigated. The simulation results demonstrate the efficacy of the proposed architecture, positioning it as a leading candidate for many vision-related tasks.