Brightness-Adjustable Low-Light Image Enhancement via Noise-Only Training Scheme.

Existing no-reference low-light image enhancement methods are trained mainly by public datasets, inevitably unable to cover most scenes in the real world. If we change the point of view, Gaussian noises reserve natural image distribution, which may be an alternative to real-world data for image enhancement. In this paper, we design a brand-new Network to brighten low-light images via a curve-based Noise-Only Training scheme (NOTNet). Specifically, our model is devoted to enhancing images to continuous brightness levels by training random noises with different means and contrasts. Extensive experiments on mainstream datasets demonstrate our model’s superiority quantitatively and qualitatively. Our code is available at https://github.com/xyn1201/NOTNet.