To address those major weaknesses of current methods of low-light enhancement, we introduce an Adaptive Semantic Prior Guided Multi-Scale Low-Light Image Enhancement (ASPG-MSLIE) framework which consists of three sections as illustrated in the abstract. In order to resolve this problem we suggest a U-Net backbone with a lightweight semantic prior extractor and a dynamic illumination integration strategy. The network is capable of maintaining the perceptual relevance of high-level semantic context, and suppressing noise in other semantic contexts, through conditioning improvement. So-dimensionality is confirmed by experiments on LOL test set (23.8 dB PSNR, 0.86 SSIM at 0.18 s inferring time) and VE-LOL (88.4% pixel accuracy), confirming state-of-the-art performance over Retinex-Net, KinD, Zero-DCE, and EnlightenGAN.
Harish et al. (Thu,) studied this question.