OPTIMASI MODEL EFFICIENTNET DAN RESNET DALAM PENGENALAN RASA NYERI KUCING BERDASARKAN DATA GAMBAR MEDIA SOSIAL X

Muhammad Ayyash, - and Indira Syawanodya, - and Yulia Retnowati, - (2025) OPTIMASI MODEL EFFICIENTNET DAN RESNET DALAM PENGENALAN RASA NYERI KUCING BERDASARKAN DATA GAMBAR MEDIA SOSIAL X. S1 thesis, Universitas Pendidikan Indonesia.

Abstract

Penelitian ini bertujuan untuk mengoptimasi model deep learning EfficientNet dan ResNet pengenal rasa nyeri pada kucing berspesies Felis catus (kucing domestik), yang berukuran seluler untuk membantu dokter hewan dalam proses diagnosis medis yang disebabkan oleh sifat inheren ketidakekspresian kucing saat merasa nyeri. Keandalan model diindikasikan oleh nilai precision dan recall yang relatif besar dan ukuran file yang lebih kecil sembari mempertahankan kemampuan inferensi yang baik. Prosesnya dilengkapi oleh dataset gambar wajah kucing yang diakuisisi dari media sosial X (sebelumnya Twitter), dengan memanfaatkan kata kunci berkaitan yang terkandung dalam tweets; kemudian di- preprocess dengan pendekatan studi semantik dan sistem penilaian nyeri kucing Feline Grimace Scale (FGS) untuk anotasi ground truth. Proses fine tuning yang dipakai pada model EfficientNet B3 dan ResNet18, melibatkan integrasi suatu jenis operasi tambahan berupa focal loss; tiga jenis optimizer, yaitu SGD, Adam, dan AdamW; dan dua teknik augmentasi data yang berbeda, tanpa dan dengan manipulasi pencahayaan gambar. Hasil yang diperoleh mengungkapkan bahwa faktor dasar yang menghambat model berkinerja lebih baik dalam penelitian ini adalah kuantitas data training yang masih kurang disebabkan oleh penyaringan data tweets yang kurang efektif, terungkap setelah perolehan ground truth menggunakan FGS. Sehingga kehilangan beberapa data yang berpotensi layak. Model terbaik diraih oleh ResNet18 dengan optimizer AdamW; mekanisme focal loss; dan augmentasi data yang tidak memanipulasi pencahayaan gambar pada testing set simulasi dunia nyata. Mencapai nilai precision terbaik, sebesar 100%, dan recall sebesar 85,71% (akurasi 92,85%). Sementara pada testing set emulasi dunia nyata, model EfficientNet B3 dengan optimizer AdamW dan augmentasi manipulasi pencahayaan gambar meraih akurasi terbesar kedua setelah Resnet18 dengan optimizer AdamW; mekanisme focal loss; dan tanpa augmentasi, sebesar 80%. Disimpulkan bahwa optimizer AdamW, teknik augmentasi data manipulasi pencahayaan gambar, dan mekanisme focal loss mampu membantu model membaik tanpa menambahkan ukuran operasi parameter-parameter dan ukuran file yang signifikan, namun harus disertai dengan model berkemampuan yang lebih besar. ---------- This study aims to optimize the EfficientNet and ResNet deep learning models for pain recognition in cats of the species Felis catus (domestic cats), which are cellular in size to assist veterinarians in the medical diagnosis process caused by the inherent nature of cat inexpressiveness in times of pain. The reliability of the model is indicated by the relatively large precision and recall values and smaller file size while maintaining good inference ability. The process is complemented by a dataset of cat face images acquired from social media X (formerly Twitter), by utilizing related keywords contained in tweets; then preprocessed with a semantic study approach and a cat pain assessment system Feline Grimace Scale (FGS) for ground truth annotation. The fine tuning process used in the EfficientNet B3 and ResNet18 models involves the integration of an additional type of operation in the form of focal loss; three types of optimizers, namely SGD, Adam, and AdamW; and two different data augmentation techniques, without and with image lighting manipulation. The results obtained revealed that the basic factor that hindered the model from performing better in this study was the quantity of training data that was still lacking due to ineffective filtering of tweets data, revealed after obtaining ground truth using FGS. So that it lost some potentially feasible data. The best model was achieved by ResNet18 with the AdamW optimizer; focal loss mechanism; and data augmentation that did not manipulate image lighting on the real-world simulation testing set. Achieving the best precision value, at 100%, and recall of 85.71% (accuracy of 92.85%). While on the real-world emulation testing set, the EfficientNet B3 model with the AdamW optimizer and image lighting manipulation augmentation achieved the second highest accuracy after Resnet18 with the AdamW optimizer; focal loss mechanism; and without augmentation, at 80%. It was concluded that the AdamW optimizer, image lighting manipulation data augmentation techniques, and focal loss mechanisms were able to help the model improve without adding significant parameter operation sizes and file sizes, but must be accompanied by a larger capable model.

	Text S_RPL_2006434_Title.pdf - Published Version Download (1MB)
	Text S_RPL_2006434_Chapter1.pdf - Published Version Download (159kB)
	Text S_RPL_2006434_Chapter2.pdf - Published Version Restricted to Staf Perpustakaan Download (1MB) | Request a copy
	Text S_RPL_2006434_Chapter3.pdf - Published Version Download (1MB)
	Text S_RPL_2006434_Chapter4.pdf - Published Version Restricted to Staf Perpustakaan Download (5MB) | Request a copy
	Text S_RPL_2006434_Chapter5.pdf - Published Version Download (135kB)
	Text S_RPL_2006434_Appendix.pdf - Published Version Restricted to Staf Perpustakaan Download (4MB) | Request a copy

Item Type:	Thesis (S1)
Uncontrolled Keywords:	computer vision, feline grimace scale, gambar media sosial, kucing, media sosial twitter, media sosial X, pengenalan rasa nyeri
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science T Technology > T Technology (General)
Divisions:	UPI Kampus cibiru > S1 Rekayasa Perangkaat Lunak
Depositing User:	Muhammad Ayyash
Date Deposited:	14 Jul 2025 08:22
Last Modified:	14 Jul 2025 08:22
URI:	http://repository.upi.edu/id/eprint/134145

Actions (login required)

View Item