Lumina: Seasonal Color Analysis System with Explainable Machine Learning
Abstract
Personal color analysis, known as Armocromia, classifies individuals into seasonal types to guide clothing recommendations, yet the practice remains largely subjective and dependent on trained consultants. We present Lumina, a lightweight and interpretable system that extracts handcrafted LAB color features from three facial regions (skin, hair, and eye) via Facer-based segmentation and classifies them using classical machine learning models trained on the Deep Armocromia benchmark. Across six classifiers and two feature configurations, a tuned XGBoost model achieves 0.552 test accuracy and a 0.537 macro F1, while an OOF-guided hill-climbing ensemble reaches 0.556 accuracy, marginally exceeding the FaRL64 deep learning baseline of 0.554 with no pre-trained neural backbone required at the classification stage. SHAP analysis on XGBoost shows that eye and hair features deliver a more informative signal than skin tone alone, a result we treat as dataset- and model-specific, given that uncontrolled illumination may distort skin LAB values independently of true undertone. Spring appears to be a systematic failure case, with recall consistently below 0.42 and misclassifications distributed across all remaining seasons. The primary contributions of this work are the first interpretable classical ML benchmark on the Deep Armocromia dataset, and SHAP-based validation of regional feature contribution. These results demonstrate that classical machine learning with principled regional feature design is a competitive and transparent alternative to deep learning for personal color analysis.
Authors: Ngwe Yee Pearl Ou, Jonathan H. Chan