Color Normalization Through a Simulated Color Checker Using Generative Adversarial Networks

Abstract

Digital cameras often struggle to reproduce the true colors perceived by the human eye due to lighting geometry and illuminant color. This research proposes an innovative approach for color normalization in digital photographs. A machine learning algorithm combined with an external physical color checker achieves color normalization. To address the limitations of relying on a physical color checker, our approach employs a generative adversarial network capable of replicating the color normalization process without the need for a physical reference. This network (GAN-CN-CC) incorporates a custom loss function specifically designed to minimize errors in color generation. The proposed algorithm yields the lowest coefficient of variation in the normalized median intensity (NMI), while maintaining a standard deviation comparable to that of conventional methods such as Gray World and Max-RGB. The algorithm eliminates the need for a color checker in color normalization, making it more practical in scenarios where inclusion of the checker is challenging. The proposed method has been fine-tuned and validated, demonstrating high effectiveness and adaptability.