<![CDATA[Facial skin tone classification plays an important role in cosmetic recommendations and digital beauty applications. Existing approaches frequently rely on single-domain feature extraction combined with conventional classifiers, which may limit representational coverage and analytical transparency. This study proposes a hybrid feature-fusion framework that integrates HSV color descriptors, GLCM statistical texture features, and LBP micro-texture patterns into a unified feature representation, which is then classified using a Light Gradient Boosting Machine (LightGBM). Experiments were conducted on 899 facial images from the Kaggle Skin Tone dataset, categorized into White, Olive, and Dark Brown classes. A stratified 70:30 hold-out scheme was applied to evaluate model performance. The results indicate that combining heterogeneous visual descriptors improves classification accuracy compared to individual feature domains. Under the selected experimental configuration, the LightGBM classifier achieved an accuracy of 89.67%, with consistent precision and recall values across classes. In addition to overall performance metrics, balanced accuracy and confusion matrix analysis were employed to examine class-level prediction behavior. SHAP-based interpretability analysis was further utilized to identify feature contribution patterns, showing that chromatic descriptors played a prominent role while texture features contributed to boundary refinement. These findings suggest that hybrid feature integration combined with boosting-based modeling provides a structured and interpretable approach for moderate-scale skin tone classification tasks.]]>
Copyrights © 2026
