CD-iNet: Deep Invertible Network for Perceptual Image Color Difference Measurement

Zhihua Wang, Keshuo Xu, Keyan Ding, Qiuping Jiang, Yifan Zuo, Zhangkai Ni, Yuming Fang

Abstract: Image color difference (CD) measurement, a crucial concept in color science and imaging technology, aims to quantify the perceived difference between two colors. Most widely recognized CD formulae are recommended by the Commission Internationale de l’Èclairage (CIE), which are tailored to homogeneous color patches and may not generalize effectively to images encompassing diverse content. Developing effective CD metrics for natural images remains an active and ongoing area of research. Drawing inspiration from the design principles found in CIE-recommended formulae, which place a premium on achieving a perceptually uniform color space, we posit that an ideal color space should adhere to the following criteria: (1) Characterizing any color pixel with three degrees of freedom, which is necessary and sufficient; (2) The visual distance between two pixels is proportional to the Euclidean distance, i.e., perceptual uniformity; (3) The transformation between color spaces is inherently reversible and has low computational complexity. To satisfy these criteria, we investigate to leverage deep invertible neural network (DINNs) to learn an invertible coordinate transform, in which the Euclidean distance is employed to compute the CD on a pixel-by-pixel basis within the transformed color space and subsequently average the resulting CD map to obtain the global CD for a pair of images. By using DINNs, the acquired coordinate transform can maintain three-dimensional properties and mathematical invertibility. The resulting metric, referred to as CD-iNet, is end-to-end optimized on color patch datasets and image datasets simultaneously. Extensive quantitative and qualitative experiments on smartphone photograph datasets demonstrate the superiority of CD-iNet over existing metrics. Besides, CD-iNet can produce competitive local CD maps without requiring dense supervision and be robust against geometric distortions. More importantly, the transformed color space exhibits reasonable characteristics of perceptual uniformity, e.g., low cross-contamination between color attributes. Codes are available at: https://github.com/hellooks/CD-iNet.