Unsupervised Low-Light Image Enhancement Algorithm Based on Illumination Adaptive Enhancement

Images captured under low-light conditions frequently exhibit degradation due to a range of factors, such as reduced contrast, noise, and color distortion. Addressing these challenges, this paper presents a novel low-light image enhancement (LIE) algorithm that draws upon Retinex theory and self-supervised learning. The proposed algorithm initializes by extracting rich adaptive priors through the constraint of reflectance maps utilizing paired low-light input images. In the preprocessing phase, the input image is directed through a projection module to eliminate unwanted features prior to decomposition. After decomposing the output feature map into reflection and illumination components, for the illumination component, an adaptive enhancement matrix module is learned to secure appropriate lighting conditions. Concurrently, a denoising module is applied to the reflectance component to eradicate noise. Ultimately, the enhanced image is produced by performing an element-wise multiplication of the outputs from the illumination and reflectance components. Experimental outcomes on public datasets indicate that the proposed method, which requires less manual crafting of priors, attains superior performance when compared to state-of-the-art (SOTA) unsupervised algorithms. Quantitative evaluations reveal that our algorithm surpasses the latest SOTA unsupervised models, with a notable improvement of approximately 6.4% in PSNR and 4.1% in SSIM.