SISTEM CERDAS DETEKSI EMOSI WAJAH REAL-TIME UNTUK EDUKASI DIGITAL MENGGUNAKAN CNN MOBILENETV3 DENGAN TRANSFER LEARNING

Arifani, Kahpi Baiquni (2026) SISTEM CERDAS DETEKSI EMOSI WAJAH REAL-TIME UNTUK EDUKASI DIGITAL MENGGUNAKAN CNN MOBILENETV3 DENGAN TRANSFER LEARNING. Masters thesis, Universitas Pembangunan Nasional Veteran Jawa Timur.

	Text (Cover) 01_Cover_hingga_Daftar_Arti_Simbol_fix.pdf Download (7MB)
	Text (Bab 1 Tesis) 02_Bab_1.pdf Download (134kB)
	Text (Bab 2 Tesis) 03_Bab_2.pdf Restricted to Repository staff only until 10 August 2029. Download (1MB) \| Request a copy
	Text (Bab 3 Tesis) 04_Bab_3.pdf Restricted to Repository staff only until 10 August 2029. Download (782kB) \| Request a copy
	Text (Bab 4 Tesis) 05_Bab_4.pdf Restricted to Repository staff only until 10 August 2029. Download (2MB) \| Request a copy
	Text (Bab 5 Tesis) 06_Bab_5.pdf Download (214kB)
	Text (Daftar Pustaka Tesis) 07_Daftar_Pustaka.pdf Download (127kB)
	Text (Lampiran Tesis) 08_Lampiran.pdf Restricted to Repository staff only until 10 August 2029. Download (229kB) \| Request a copy

Abstract

Non-intrusive affective monitoring of students has become an important requirement for Indonesian digital education platforms, yet the deployment of Facial Expression Recognition (FER) faces two major challenges: limited hardware capacity in school environments and the domain gap between international benchmark data and the local population. This research proposes a two-stage scheme to address both challenges separately. Stage I applies domain-matched Knowledge Distillation (KD), training MobileNetV3 student models (Large and Small) under the guidance of a frozen DDAMFN++ teacher model on the identical RAF-DB domain. Stage II applies discriminative fine-tuning from the Stage I checkpoint to the Indonesian hybrid IMED-IMSFD dataset as the target domain. Stage I results show the MobileNetV3-Large student surpassing its teacher on Balanced Accuracy (73.16% versus 71.30%), while a comparative study demonstrates that KD effectiveness is highly dependent on teacher-student domain alignment: direct cross-domain distillation degrades student performance, whereas domain-matched distillation is beneficial. An in-domain vs cross-domain test on all four models, using an identical protocol and sample size (N=30/class) on both sides, confirms a sharp domain gap (a 39.5-65.7 percentage-point drop outside each model's training domain), largest for the Stage II student despite its highest in-domain performance, validating the need for explicit domain adaptation. Stage II results show Balanced Accuracy rising from a 33.44% zero-shot baseline to 91.89% after fine-tuning (+58.45 points) on the IMED-IMSFD target domain. Post-Training Quantization (PTQ) to INT8 on the Stage I checkpoint achieves up to 3.58x compression with a modest accuracy drop for the Large variant (-6.27 points Balanced Accuracy), but a drastic collapse for the Small variant (-45.57 points), so only MobileNetV3-Large is recommended for edge deployment. The same quantization on the Stage II checkpoint retains 3.59x compression with a 7.91-point Balanced Accuracy drop (91.89% to 83.98%), confirming MobileNetV3-Large remains deployment-viable on the target domain. Structural/functional integration into the 2GT Smart Classroom platform, including dynamically selectable quantized models from both stages, was successfully verified. This research concludes that the two-stage scheme, domain-matched KD followed by transfer learning, is an effective approach for building compact FER models relevant to the Indonesian digital education context. Keywords: Facial Expression Recognition, Knowledge Distillation, Transfer Learning, MobileNetV3, Digital Education

Item Type:

Thesis (Masters)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Diyasa, I Gede Susrama Mas	NIDN0019067008	igusrama.if@upnjatim.ac.id
Thesis advisor	Agussalim, Agussalim	NIDN0911088501	agussalim.si@upnjatim.ac.id

Subjects:

T Technology > T Technology (General)
T Technology > T Technology (General) > T58.6-58.62 Management Information Systems

Divisions:

Faculty of Computer Science > Magister Information Technology

Depositing User:

Mr. Kahpi Baiquni Arifani

Date Deposited:

10 Aug 2026 08:47

Last Modified:

10 Aug 2026 08:47

URI:

https://repository.upnjatim.ac.id/id/eprint/58517