OPTIMASI PADA TRANSFER LEARNING UNTUK KLASIFIKASI EMOSI WAJAH PADA DATASET FER-2013

Nida Muhliya Barkah, NIM.: 22206052003 (2024) OPTIMASI PADA TRANSFER LEARNING UNTUK KLASIFIKASI EMOSI WAJAH PADA DATASET FER-2013. Masters thesis, UIN SUNAN KALIJAGA YOGYAKARTA.

Preview	Text (OPTIMASI PADA TRANSFER LEARNING UNTUK KLASIFIKASI EMOSI WAJAH PADA DATASET FER-2013) 22206052003_BAB-I_IV-atau-V_DAFTAR-PUSTAKA.pdf - Published Version Download (2MB) | Preview
	Text (OPTIMASI PADA TRANSFER LEARNING UNTUK KLASIFIKASI EMOSI WAJAH PADA DATASET FER-2013) 22206052003_BAB-II_sampai_SEBELUM-BAB-TERAKHIR.pdf - Published Version Restricted to Registered users only Download (6MB) | Request a copy

Abstract

Facial expressions are essential in non-verbal communication, as they naturally convey human emotions during personal interactions. Emotion recognition from facial expressions through computer vision is a popular topic in affective computing. The swift advancement of deep learning has resulted in its growing application in facial emotion recognition. However, it still encounters a significant challenge due to the necessity for a substantial amount of data to operate effectively. Numerous studies have utilized transfer learning to tackle this problem, yet a standardized approach for implementing transfer learning in facial emotion recognition has not been established. This research classifies facial emotion images into seven categories using three transfer learning models: ResNet-50, Xception, and Inception V3 on the FER-2013 dataset. Various experiments, including data pre-processing, hyperparameter tuning, and model training techniques, have been conducted. The data pre-processing results show that each model requires different input image sizes to achieve the best accuracy. Hyperparameter tuning results indicate accuracy improvements of 6.3527% for ResNet-50, 4.6949% for Inception V3, and 1.039% for Xception. Augmentation experiments show that augmenting only the disgust class yields better accuracy than augmenting all classes. Model training experiments reveal that the freeze fine-tuning method is not better compared to the fine-tuning method on datasets with thousands of samples but is superior to the freeze layer method commonly used in many studies. The best accuracy values for each model are 64.8927% for ResNet-50, 65.8261% for Xception, and 66.3973% for Inception V3.

Share this knowledge with your friends :

Actions (login required)

View Item