<![CDATA[Manual classification of facial skin types remains subjective, time-consuming, and inconsistent, complicating appropriate skincare selection and increasing risks of irritation or acne. This research developed an automatic facial skin type detection system (acne, dry, normal, oily) using a hybrid fine-tuned MobileNetV2 and RBF-SVM approach. A dataset of 2,000 images from Roboflow underwent intensive augmentation, deep fine-tuning of MobileNetV2, 1280-dimensional feature extraction, StandardScaler normalization, and RBF-SVM training with optimal hyperparameters (C = 30, gamma = 0.001) obtained via 5-fold GridSearchCV. The system incorporates Haar Cascade face detection for real-world robustness. The final hybrid model achieved the highest test accuracy of 90.31% (macro F1-score 0.90) on an independent 196-image test set, outperforming the non-fine-tuned baseline by 15.31 points and the pure CNN-softmax variant by 4.95 points. The entire pipeline has been successfully implemented as a public web application named FACEDX using React.js with Tailwind CSS (frontend) and Flask with Gunicorn (backend), deployed on Railway with a custom domain and comprehensive eror handling. This application can be utilized by the general public, dermatologists, and the cosmetic industry for more accurate and personalized skincare recommendations.]]>
Copyrights © 2026
