Deep Learning-Based Skin Type Identification and Personalized Skincare Recommendations
S. A. D. S. Sanjana *
Department of Information and Communication Technology, Faculty of Humanities and Social Sciences, University of Sri Jayewardenepura, Nugegoda, Sri Lanka.
D. P. M. Perera
Department of Information and Communication Technology, Faculty of Humanities and Social Sciences, University of Sri Jayewardenepura, Nugegoda, Sri Lanka.
*Author to whom correspondence should be addressed.
Abstract
The skin is the outermost layer of the body and is crucial for women's beauty and health. Understanding one's skin type is essential for effective skincare, as a lack of knowledge can lead to adverse effects. This study proposes a deep learning-based system designed to identify women's facial skin types and provide personalized skincare recommendations. Leveraging the MobileNetV2 architecture as the base model, a convolutional neural network, the model was trained to classify five skin types: Dry, Oily, Normal, Combination, and Sensitive. Data were collected from cosmetology specialists and online datasets, with preprocessing and image augmentation to enhance the dataset. The approach involved modifying the MobileNetV2 by adding GlobalAveragePooling2D and custom dense layers tailored for skin type classification. The developed model achieved a testing accuracy of 93.78% and was integrated into a Flask-based web application. This application allows users to upload facial images, provide their age and allergy status, and receive tailored skincare routines based on the predicted skin type. User testing demonstrated a high level of satisfaction, with 64% of respondents finding the skin type predictions accurate and 61% considering the skincare recommendations useful. The website's interface received positive feedback, with 95% of respondents rating it as user-friendly. While the model performed well in classifying most skin types, it encountered difficulties in differentiating between visually similar types, such as Combination and Sensitive, where misclassifications occurred due to overlapping visual features. Additionally, the study faced limitations stemming from image quality variability, and the constraints of rule-based recommendations. Furthermore, the current dataset did not fully represent all demographic groups and diverse skin tones. Overall, this study contributes to the scientific community by presenting a practical, real-time application of AI in dermatology that bridges the gap between technology and everyday skincare needs.
Keywords: Skin type identification, Skincare routine, MobileNetv2, convolutional neural networks, deep learning