Priori Knowledge Makes Low-Light Image Enhancement More Reasonable

This paper presents a priori knowledge-based low-light image enhancement framework, termed Priori DCE (Priori Deep Curve Estimation). The priori knowledge consists of two key aspects: 1) Enhancing low-light image is an ill-posed task, as the brightness of the enhanced image corresponding to a low-light image is uncertain. To resolve this issue, we incorporate priori channels into the model to guide the brightness of the enhanced image; 2) During the enhancement of low-light image, the brightness of pixel may increase or decrease. This paper refer to the probability of a pixel’s brightness increasing/decreasing as its prior enhancement/suppression probability. Intuitively, pixel with higher brightness should have a higher priori suppression probability, while pixel with lower brightness should have a higher priori enhancement probability. Inspired by this, we propose a enhancement function that adaptively adjusts the priori enhancement probability based on variation in pixel brightness. In addition, the enhanced images generated by existing methods often suffer from varying degrees of noise, blur, and artifacts. This is primarily because each pixel in the enhanced image are typically computed from the corresponding pixel at the same position in the low-light image and a limited number of neighboring pixels, which fails to account for the global visual balance during the enhancement process. To address this issue, we propose the Global-Attention Block (GA Block). The GA Block ensures that, during the low-light image enhancement process, each pixel in the enhanced image is computed based on all the pixels in the low-light image. This approach facilitates interactions between all pixels in the enhanced image, thereby achieving visual balance. The experimental results on the LOLv2-Synthetic dataset demonstrate that Priori DCE has a significant advantage. Specifically, compared to the SOTA Retinexformer, the Priori DCE improves PSNR index and SSIM index from 25.67 and 92.82 to 29.49 and 93.6, respectively, while NIQE index decreases from 3.94 to 3.91. The code and experiments will soon be open-sourced at https://github.com/zefeichen/PrioriDCE.