Illumination Removal for Accurate Skin Color Extraction Using Dermoscopy and Transformer-Based Modeling

Skin conditions such as vitiligo, scars, melasma, and age spots impact not only cosmetic appearance but also an individual's emotional and psychological well-being.Many patients attempt to conceal these conditions using cosmetic products; however, accurately determining intrinsic skin color is challenged by significant variation in perceived color under different lighting conditions.The dependency on ambient lighting introduces inconsistencies, making reliable assessment of the true skin tone difficult.Prior large-scale patient studies document the substantial psychosocial burden associated with visible pigmentary disorders (e.g., vitiligo), and systematic reviews have similarly demonstrated the negative impact of conditions such as melasma on quality of life 1-3 .Moreover, technical and perceptual studies show that ambient and background illumination materially alter perceived skin color and complicate instrument-or camera-based color assessment, motivating the need for illumination-robust measurement methods 4,5 .To address the challenge of inconsistent skin color perception caused by ambient lighting, we propose an illumination removal system combining dermoscopy imaging and Transformer-based deep learning modeling.Dermoscopy ensures stable imaging conditions, while our artificial intelligence model further corrects residual illumination variations.All dermoscopic images were acquired using a single smartphone-dermoscope combination (Galaxy S23 equipped with the ILLUCO IDS-1100 dermoscope).All images were obtained using contact-type polarized dermoscopy.Automatic exposure and white-balance settings were fixed throughout image acquisition to minimize device-induced variability.The training architecture is illustrated in Fig. 1A, and the overall workflow in Fig. 1B.Despite dermoscopic control, minor illumination inconsistencies persist.We developed a Transformer model that takes small patches (2020 pixels) as input to map unevenly illuminated RGB values to true values.The model outputs corrected pixel values, which are averaged to yield the estimated skin color-thus enabling consistent extraction under pixel-level lighting variations.For dataset construction, we used 195 standardized color kits with predefined RGB values.Each image was center-cropped to 1,0001,000 pixels, then segmented by a sliding window of 2020 pixels, generating 2,500 samples per kit (487,500 total).