Electronic Theses and Dissertation

Penelitian ini bertujuan untuk membandingkan performa dua arsitektur deep learning, resnet50 dan efficientnetv2b0, dalam klasifikasi emosi wajah berbasis citra, serta mengembangkan dataset baru bernama nafed-7 sebagai solusi atas keterbatasan dataset sebelumnya dan mendukung penelitian lanjutan dalam klasifikasi emosi berbasis citra. dataset nafed-7 dikumpulkan dari 100 individu, dipra-proses dengan deteksi wajah menggunakanmediapipe, resize ke 224×224 piksel, serta augmentasi data. model resnet50 dan efficientnetv2b0 dilatih menggunakan pendekatan transfer learning berbasis imagenet, dengan total 16 kombinasi hyperparameter. evaluasi dilakukan menggunakan metrik accuracy, precision, recall, f1-score, dan confusion matrix. hasil pengujian menunjukkan bahwa resnet50 mencapai akurasi 45%, presisi 50%, recall 45%, dan f1-score 44% pada konfigurasi terbaik, dengan waktu pelatihan 4 menit 22 detik. sementara itu, efficientnetv2b0 hanya mencapai 32% di semua metrik dan memerlukan waktu pelatihan lebih lama. dengan hasil tersebut, resnet50 dinilai lebih unggul, efisien, dan konsisten, serta dipilih sebagai model utama untuk sistem deteksi emosi berbasis web. penelitian selanjutnya disarankan untuk memperluas data dan meningkatkan pendekatan pelatihan guna memperoleh model yang lebih kuat dan akurat di berbagai kondisi nyata. kata kunci : klasifikasi emosi, ekspresi wajah, deep learning, resnet50, efficientnetv2b0, transfer learning, nafed-7

This study aims to compare the performance of two deep learning architectures, ResNet50 and EfficientNetV2B0, in image-based facial emotion classification, as well as to develop a new dataset called NaFED-7 as a solution to the limitations of previous datasets and to support further research in image-based emotion classification. The NaFED-7 dataset was collected from 100 individuals, pre-processed with face detection using MediaPipe, resized to 224×224 pixels, and data augmentation. The ResNet50 and EfficientNetV2B0 models were trained using a transfer learning approach based on ImageNet, with a total of 16 hyperparameter combinations. Evaluation was conducted using accuracy, precision, recall, F1-score, and confusion matrix metrics. Test results show that ResNet50 achieves 45% accuracy, 50% precision, 45% recall, and 44% F1-score at the best configuration, with a training time of 4 minutes and 22 seconds. Meanwhile, EfficientNetV2B0 only achieved 32% across all metrics and required a longer training time. With these results, ResNet50 is deemed superior, efficient, and consistent, and selected as the primary model for the web-based emotion detection system. Further research is recommended to expand the data and improve the training approach to obtain a stronger and more accurate model under various real-world conditions. Keywords: Emotion Classification, Facial Expressions, Deep Learning, ResNet50, EfficientNetV2B0, Transfer Learning, NaFED-7.