SCNet: A Self-Calibrating Unsupervised Low-Light Image Enhancement Network
Runze Zhang, Shuanglong Yao, Liang Lü, Xing Wang
Abstract
Image acquired by camera sensors suffers from low contrast and poor visibility under low-light conditions. Given the importance of enhancing low-light images, this article presents an innovative method known as the self-calibrating neural network (SCNet). This approach involves the extraction of a lighting adjustment feature map (LafMap) from the input image. Matrix operations are then employed on the input image to derive an intermediate image, which is subsequently optimized through self-calibration (SC) to produce the final output image. The SCNet approach has proven effective in experimentation, as indicated by the peak signal-to-noise ratio (PSNR) values, demonstrating a 6.1% increase compared with the baseline on the low-light dataset (LOL dataset). The SCNet can be trained without the necessity for reference images, overcoming the challenge of insufficient enhancement often observed in very low-light images, which is a prominent problem in existing unsupervised low-light enhancement methods.