Unsupervised Low-Light Image Enhancement Algorithm Based on Prior Information

Images captured in dimly illuminated surroundings often suffer from low contrast and severe loss of details, which directly affect the accuracy of subsequent image classification, recognition, and detection tasks. This paper addresses the challenges of low-light image enhancement, which relies on real data training and the insufficient availability of effective information in low-light images. An unsupervised low-light image enhancement algorithm, based on prior information, is proposed by us. By performing histogram equalization on the preprocessed images before network training, hidden information in the images is obtained. The optimization of initialization information is achieved by extracting the reflectance and illumination maps of the low-light images, which preserves the relative structural integrity of the images and improves the brightness restoration effect. The experimental outcomes validate the efficacy of the proposed algorithm in restoring brightness and reproducing natural colors in images. In comparison to other algorithms, it demonstrates superior performance on the LOL dataset in terms of peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and natural image quality evaluation metrics (NIQE). Specifically, it achieves improvements of 1.433 dB, 0.040, and 1.285, respectively.