LLMEV: Low Light Image Restoration with Multiple Exposures and Visibility Levels

Abstract

Image taken in difficult lighting situations from dim settings to extremely dark scenes still are a recurring problem in modern photography. Specifically, this research addresses the complex problem of low-light picture restoration, including overexposure in bright regions, and inadequate dark image restoration. Using a Generative Adversarial Network (GAN) model to address all lighting-complexity-related problems and produce an overall solution is the primary aim. The method used is training a GAN model to understand and correct specifics related to varying degrees of low-light conditions. The model is trained on a wide range of data that includes images with varying brightness across specific regions, extremely dark images, and slightly dim images. Through adversarial training, the GAN can find the best enhancements so that brilliant areas are recovered without being overexposed, poorly lit sections are brought back into suitable visibility, and very dark areas are properly illuminated. Following 100 epochs of training on lighter datasets, each comprising fewer than 600 image pairs, the LLMEV model demonstrates the ability to restore low-light images, achieving mean output scores of 21.58 dB for Peak Signal-to-Noise Ratio (PSNR) and 0.885 for Structural Similarity Index (SSIM).