Electronic Theses and Dissertation

Penelitian ini bertujuan untuk mengklasifikasikan jenis kulit wajah manusia menggunakan pendekatan computer vision berbasis deep learning dengan arsitektur EfficientNet-B0 dan Vision Transformer (ViT). Dataset yang digunakan berjumlah 9.050 citra yang terdiri dari data pelatihan, validasi, dan pengujian, dengan 905 citra digunakan sebagai data uji. Data diklasifikasikan ke dalam empat kategori, yaitu kulit berjerawat, berminyak, kering, dan normal. Tahap pra-pemrosesan meliputi pembersihan data, penyesuaian ukuran citra menjadi 224 × 224 piksel, konversi ke format RGB, serta normalisasi menggunakan parameter ImageNet. Model dilatih menggunakan pendekatan transfer learning, data augmentation, serta strategi pelatihan dua tahap yaitu feature extraction dan fine-tuning untuk meningkatkan performa klasifikasi. Hasil pengujian menunjukkan bahwa model Vision Transformer (ViT) memperoleh performa terbaik dengan nilai accuracy sebesar 88,62%, precision sebesar 88,64%, recall sebesar 88,62%, dan F1-score sebesar 88,59%, sedangkan EfficientNet-B0 memperoleh accuracy sebesar 71,49%, precision sebesar 74,82%, recall sebesar 71,49%, dan F1-score sebesar 71,41%. Model Vision Transformer (ViT) menunjukkan kemampuan yang lebih baik dalam mengklasifikasikan citra secara akurat dan seimbang dibandingkan EfficientNet-B0. Hasil penelitian ini menunjukkan bahwa pendekatan deep learning berbasis transformer lebih efektif dalam menangani klasifikasi citra jenis kulit wajah manusia.

Kata kunci: computer vision, klasifikasi citra, jenis kulit wajah, EfficientNet-B0, Vision Transformer, deep learning.

This study aims to classify human facial skin types using a deep learning-based computer vision approach with EfficientNet-B0 and Vision Transformer (ViT) architectures. The dataset consists of 9,050 images divided into training, validation, and testing sets, with 905 images used as test data. The data is classified into four categories: acne-prone, oily, dry, and normal skin. The preprocessing stage includes data cleaning, resizing images to 224 × 224 pixels, converting them to RGB format, and normalization using ImageNet parameters. The models were trained using a transfer learning approach, data augmentation, and a two-stage training strategy consisting of feature extraction and fine-tuning to improve classification performance. The test results show that the Vision Transformer (ViT) achieved the best performance with an accuracy of 88.62%, precision of 88.64%, recall of 88.62%, and F1-score of 88.59%, while EfficientNet-B0 achieved an accuracy of 71.49%, precision of 74.82%, recall of 71.49%, and F1-score of 71.41%. The Vision Transformer (ViT) demonstrates superior capability in performing accurate and balanced image classification compared to EfficientNet-B0. The results indicate that transformer-based deep learning approaches are more effective in handling facial skin type classification tasks. Keywords: computer vision, image classification, facial skin types, EfficientNet-B0, Vision Transformer, deep learning.