Learning Image-adaptive 3D Lookup Tables for High Performance Photo Enhancement in Real-time

Recent years have witnessed the increasing popularity of learning basedmethods to enhance the color and tone of photos. However, many existing photoenhancement methods either deliver unsatisfactory results or consume too muchcomputational and memory resources, hindering their application tohigh-resolution images (usually with more than 12 megapixels) in practice. Inthis paper, we learn image-adaptive 3-dimensional lookup tables (3D LUTs) toachieve fast and robust photo enhancement. 3D LUTs are widely used formanipulating color and tone of photos, but they are usually manually tuned andfixed in camera imaging pipeline or photo editing tools. We, for the first timeto our best knowledge, propose to learn 3D LUTs from annotated data usingpairwise or unpaired learning. More importantly, our learned 3D LUT isimage-adaptive for flexible photo enhancement. We learn multiple basis 3D LUTsand a small convolutional neural network (CNN) simultaneously in an end-to-endmanner. The small CNN works on the down-sampled version of the input image topredict content-dependent weights to fuse the multiple basis 3D LUTs into animage-adaptive one, which is employed to transform the color and tone of sourceimages efficiently. Our model contains less than 600K parameters and takes lessthan 2 ms to process an image of 4K resolution using one Titan RTX GPU. Whilebeing highly efficient, our model also outperforms the state-of-the-art photoenhancement methods by a large margin in terms of PSNR, SSIM and a colordifference metric on two publically available benchmark datasets.