Statistical image processing (Decorrelation Stretch) and deep learning (CycleGANs) to restore images of faded artworks

Cultural properties gradually fade and deteriorate over time due to various factors, including exposure to ultraviolet and infrared radiation, fluctuations in temperature and humidity that cause the degradation of organic materials, and the accumulation of dust and other contaminants on the surface. These factors diminish the visibility of the artwork, often requiring professional restoration techniques that are time-consuming, costly, and potentially irreversible. This study proposes a method to digitally restore the visual clarity of cultural properties by combining statistical image processing with deep learning. First, Decorrelation Stretch (DStretch) is applied to transform the color profile of photographed images, enhancing the visibility of faded motifs such as characters and illustrations. Next, Cycle-Consistent Generative Adversarial Networks (CycleGANs) are applied to learn the color domain transformation between the DStretch-processed (faded) images and reference images of similar, minimally faded artworks. This enables the generation of estimated prefaded images without requiring any physical intervention on the actual cultural property. The proposed method facilitates digital restoration and utilization of cultural properties, contributing to their preservation and broader public access. Furthermore, by reusing and reanalyzing processed image data, this approach fosters reproducibility in cultural heritage research, which is often susceptible to subjectivity.