Analisis kinerja Arsitektur VGG19 dan ResNet101 untuk klasifikasi etnisitas berdasarkan citra wajah

Hamza, Aria Octavian (2026) Analisis kinerja Arsitektur VGG19 dan ResNet101 untuk klasifikasi etnisitas berdasarkan citra wajah. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

Preview	Text 1_cover.pdf Download (98kB) | Preview
Preview	Text 2_abstrak.pdf Download (180kB) | Preview
Preview	Text 3_skbebasplagiarism.pdf Download (893kB) | Preview
Preview	Text 4_daftarisi.pdf Download (348kB) | Preview
Preview	Text 5_bab1.pdf Download (284kB) | Preview
	Text 6_bab2.pdf Restricted to Registered users only Download (439kB) | Request a copy
	Text 7_bab3.pdf Restricted to Registered users only Download (403kB) | Request a copy
	Text 8_bab4.pdf Restricted to Registered users only Download (6MB) | Request a copy
	Text 9_bab5.pdf Restricted to Registered users only Download (190kB) | Request a copy
	Text 10_daftarpustaka.pdf Restricted to Registered users only Download (163kB) | Request a copy
	Text 11_lampiran.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

INDONESIA; Klasifikasi etnisitas berbasis citra wajah merupakan salah satu tugas dalam bidang computer vision yang memiliki tingkat kompleksitas tinggi akibat kemiripan fitur visual antar kelompok etnis yang sering kali bersifat halus dan saling tumpang tindih. Penelitian ini bertujuan untuk menganalisis dan membandingkan kinerja dua arsitektur Convolutional Neural Network (CNN), yaitu VGG19 dan ResNet101, dalam tugas klasifikasi etnisitas berbasis citra wajah menggunakan dataset FairFace yang mencakup enam kategori etnisitas: Caucasian, Negroid, East Asian, Southeast Asian, India, dan Middle Eastern. Metodologi penelitian mengikuti kerangka CRISP-DM dengan skema pelatihan yang seragam, meliputi strategi transfer learning, partial freezing, dan kepala klasifikasi kustom yang identik pada kedua arsitektur. Untuk menangani ketidakseimbangan distribusi kelas, tiga metode penyeimbangan data dievaluasi, yaitu oversampling dengan augmentasi agresif, undersampling, dan class weight. Hasil eksperimen menunjukkan bahwa metode oversampling menghasilkan performa terbaik pada kedua model. Pada kondisi tersebut, VGG19 mencapai akurasi test set sebesar 72,87% dan mengungguli ResNet101 yang memperoleh akurasi 69,83%. Dari sisi efisiensi komputasi, ResNet101 membutuhkan waktu pelatihan lebih singkat dibandingkan VGG19. Berdasarkan hasil evaluasi, VGG19 terbukti lebih unggul dibandingkan ResNet101 dengan perolehan akurasi test set 72,87%. ENGLISH: Ethnicity classification based on facial images is a highly complex task in the field of computer vision, owing to the subtle and often overlapping visual features among different ethnic groups. This study aims to analyze and compare the performance of two Convolutional Neural Network (CNN) architectures, VGG19 and ResNet101, in facial image-based ethnicity classification using the FairFace dataset, which encompasses six ethnicity categories: Caucasian, Negroid, East Asian, Southeast Asian, Indian, and Middle Eastern. The research methodology follows the CRISP-DM framework with a uniform training scheme, including transfer learning strategies, partial layer freezing, and identical custom classification heads applied to both architectures. To address class distribution imbalance, three data balancing methods were evaluated: oversampling with aggressive augmentation, undersampling, and class weighting. Experimental results demonstrate that the oversampling method consistently yielded the best performance across both models. Under this condition, VGG19 achieved a test set accuracy of 72,87%, outperforming ResNet101 which obtained 69.83%. In terms of computational efficiency, ResNet101 required nearly half the training time compared to VGG19. Based on the evaluation results, VGG19 outperformed ResNet101, achieving a test set accuracy of 72,87% and demonstrating more consistent performance.

Actions (login required)

View Item