Efficient Image Enhancement via Representative Color Transform

We propose an improved representative color transformation (RCT++), which is an effective framework to describe complex color transformations between low- and high-quality images. We identify the representative colors and features of the input image. For each representative color, we estimate a transformed color that represents its enhanced version. Then, we enhance all input colors by interpolation, taking into account the similarity between input pixels and representative features. We further improve the original RCT framework by introducing the reconstruction term, which clarifies the representative colors, and the entropy term, which diversifies the representative features. Finally, we develop the enhancement network to achieve fast and lightweight image enhancement. Comprehensive experiments on various image enhancement tasks validate our superiority in both effectiveness and efficiency. Our method exceeds recent state-of-the-art methods in efficient image enhancement on MIT-Adobe 5K, Low Light, and Underwater Image Enhancement Benchmark datasets, with comparable computational and memory costs.