Landmark-Centric Perceptual Skin-Tone Classification and Luminance-Based Melanin Estimation for Imaging-Guided Red-Light Therapy
Author(s): Ravi Dayani, Ridham Varsani, Manjari Khatri
Publication #: 2601024
Date of Publication: 28.01.2026
Country: United States
Pages: 1-11
Published In: Volume 12 Issue 1 January-2026
DOI: https://doi.org/10.62970/IJIRCT.v12.i1.2601024
Abstract
Accurate assessment of facial skin tone and melanin is essential for tailoring red‑light therapy (RLT) protocols, yet selfie images collected in real settings are affected by illumination changes, shadows, pose, and occlusions that undermine conventional whole‑face segmentation. We present a landmark‑centric pipeline that localizes four stable facial regions—forehead, right cheek, left cheek, and sub‑labial—using a heatmap decoder with a DSNT coordinate head and an EfficientNet‑B0 backbone [1], [2], [13]. Each region is cropped and processed with dual color‑space masking (HSV + Y′CbCr) to isolate skin pixels, leveraging hexcone transforms and studio luminance/chrominance conventions [6], [7], [8], [14]. Dominant chromaticities are summarized via K‑Means and compared to a curated palette in CIE Lab using CIEDE2000 perceptual differences; voting across clusters yields per‑landmark tone labels which are aggregated to an image‑level category [3], [5]. For pigmentation, we compute a simple, device‑friendly melanin proxy from the Y′ (luminance) channel of Y′CbCr per landmark and average across regions to stabilize against local shadows and highlights [6], [8], [14]. Inference is further hardened by test‑time augmentation (original + horizontal flip) with cheek‑swap correction and coordinate remapping [10], [11]. On a diverse 1,000‑image dataset (White: 350; Brown: 400; Dark: 250), the classifier achieves 87% accuracy, with most errors confined to adjacent categories—consistent with the continuous nature of skin pigmentation and lighting effects. The approach is computationally lightweight, reproducible with standard libraries, and integrates cleanly with imaging‑guided RLT systems to enable practical dose optimization [8], [9].
Keywords: Red‑light therapy (RLT), facial landmarks, DSNT, EfficientNet‑B0, HSV/HSL, YCbCr, CIE Lab, CIEDE2000, K‑Means, test‑time augmentation, melanin proxy, skin‑tone classification.
Download/View Count: 29
Share this Article