Facial Skin Type Classification Using CNN with GLCM-Based Texture Features: A Comparative Study of MobileNet and EfficientNet-B0

Abstract

Facial skin type classification is important in dermatology and cosmetology because it supports appropriate skin care treatment selection. Common facial skin types include normal, dry, and oily skin, each having distinct texture characteristics. This study analyzes the performance of Convolutional Neural Network (CNN) models, namely MobileNet and EfficientNet-B0, for facial skin type classification and evaluates the impact of Gray Level Co-occurrence Matrix (GLCM) texture feature extraction on both models. A transfer learning approach with ImageNet pretrained weights was applied by freezing the base model while training only the classification layers. The dataset consisted of 489 facial images across three classes. The experiments were conducted using batch sizes of 8 and 16 with 20 and 30 training epochs. Model performance was evaluated using training and testing accuracy. The experimental results show that the baseline MobileNet and EfficientNet-B0 models achieved testing accuracies of 90% and 92%, respectively. After integrating GLCM texture features, both models achieved the highest testing accuracy of 93%. The incorporation of GLCM improved the testing accuracy of MobileNet by 3 percentage points and EfficientNet-B0 by 1 percentage point. These findings indicate that GLCM-based texture features can complement CNN feature extraction and improve facial skin type classification performance, although the magnitude of the improvement depends on the CNN architecture used.