Normalize-and-Realign: Controllable Image Enhancement across Multiple Exposure Levels

The objective of image enhancement is to improve the visual quality of images. It is an ill-posed task that inherently allows for multiple enhanced versions to be obtained from a given image. However, prevailing image enhancement methods yield a deterministic result for each low-quality input, disregarding different preferences and requirements. To address this issue, we propose a Reference-guided Image Enhancement Network (RIENet), which employs a normalize-and-realign paradigm for achieving multi-exposure image enhancement with precise control over brightness and color. Specifically, RIENet comprises two subnets: an Exposure Normalizing Subnet for normalizing inputs, and a reference-guided Image Realigning Subnet for realigning the normalized images according to reference images. By assigning different reference images, our method is capable of producing customized enhanced images with precise control over brightness and color faithful to the reference image. Extensive experiments demonstrate that the proposed method achieves state-of-the-art in both qualitative and quantitative results, and is able to generate training data for other low-level and high-level vision tasks. Furthermore, the potential of RIENet for video enhancement is also exhibited. To the best of our knowledge, RIENet is the first attempt to introduce reference images as guidance for achieving controllable image enhancement across multiple exposure levels. The code is available at https://github.com/FVL2020/RIENet.