PERBANDINGAN KINERJA BACKBONE CNN RESNET-50 DAN RESNET-101 UNTUK DETEKSI OBJEK BERKAMUFLASE

Dimas Aditya Permana, - (2024) PERBANDINGAN KINERJA BACKBONE CNN RESNET-50 DAN RESNET-101 UNTUK DETEKSI OBJEK BERKAMUFLASE. S1 thesis, Universitas Pendidikan Indonesia.

	Text S_RPL_2000746_Title.pdf Download (485kB)
	Text S_RPL_2000746_Chapter 1.pdf Download (238kB)
	Text S_RPL_2000746_Chapter 2.pdf Restricted to Staf Perpustakaan Download (664kB)
	Text S_RPL_2000746_Chapter 3.pdf Download (1MB)
	Text S_RPL_2000746_Chapter 4.pdf Restricted to Staf Perpustakaan Download (1MB)
	Text S_RPL_2000746_Chapter 5.pdf Download (218kB)
	Text S_RPL_2000746_Appendix.pdf Restricted to Staf Perpustakaan Download (834kB)

Abstract

Deteksi Objek Berkamuflase atau Camouflaged Object Detection (COD) adalah bidang dalam computer vision yang berfokus pada pendeteksian objek yang sulit dibedakan dari latar belakangnya. COD penting dalam studi perilaku hewan, ekologi, dan deteksi hewan terancam punah di alam liar. Dalam beberapa penelitian mengenai COD, ResNet-50 sering digunakan sebagai backbone, namun dinilai kurang efektif dalam mendeteksi fitur kompleks dan detail halus dibandingkan dengan ResNet-101 yang merupakan hal penting dalam penelitian deteksi objek berkamuflase. Oleh karena itu, penelitian ini membandingkan kinerja ResNet-50 dan ResNet-101 sebagai backbone dalam model SINet untuk deteksi objek berkamuflase untuk mengetahui backbone yang sesuai untuk penelitian terkait COD. Tiga aspek utama diuji dalam penelitian ini: kualitas segmentasi objek dibandingkan dengan ground-truth, akurasi dalam mengidentifikasi lokasi objek, serta performa keseluruhan dalam precision dan recall. Penelitian dilakukan dengan mengganti backbone SINet dari ResNet-50 ke ResNet-101, menggunakan tiga dataset utama: COD10K, CAMO, dan CHAMELEON. Model dilatih dalam dua fase: pertama selama 40 epoch dan kedua dengan early stopping, 28 epoch untuk ResNet-101 dan 38 epoch untuk ResNet-50. Evaluasi kinerja dilakukan menggunakan metrik S-Measure, E-Measure, MAE, dan Weighted F-Measure. Hasil menunjukkan bahwa ResNet-101 memiliki performa lebih baik dibandingkan ResNet-50 dalam akurasi, segmentasi, precision, dan recall. ResNet-101 mencapai 88.80% dalam S-Measure pada dataset CHAMELEON, sedangkan ResNet-50 mencapai 87.80%. Dalam Weighted F-Measure, ResNet-101 menunjukkan peningkatan performa 2-3% dibandingkan ResNet-50, terutama pada dataset CAMO. Teknik early stopping efektif mengurangi overfitting tanpa mengorbankan akurasi, terutama pada ResNet-101. -------- Camouflaged Object Detection (COD) is a field in computer vision focused on detecting objects that are difficult to distinguish from their backgrounds. COD is important in the study of animal behavior, ecology, and the detection of endangered species in the wild. In several studies on COD, ResNet-50 is often used as the backbone, but it is considered less effective in detecting complex features and fine details compared to ResNet-101, which is crucial in camouflage object detection research. Therefore, this study compares the performance of ResNet-50 and ResNet-101 as backbones in the SINet model for camouflaged object detection to determine the most suitable backbone for COD-related research. Three key aspects are examined in this study: the quality of object segmentation compared to the ground truth, the accuracy in identifying object locations, and overall performance in terms of precision and recall. The study is conducted by replacing the SINet backbone from ResNet-50 to ResNet-101, using three main datasets: COD10K, CAMO, and CHAMELEON. The model is trained in two phases: the first for 40 epochs and the second with early stopping, 28 epochs for ResNet-101 and 38 epochs for ResNet-50. Performance evaluation is conducted using S-Measure, E-Measure, MAE, and Weighted F-Measure metrics. The results show that ResNet-101 outperforms ResNet-50 in accuracy, segmentation, precision, and recall. ResNet-101 achieved 88.80% in S-Measure on the CHAMELEON dataset, while ResNet-50 reached 87.80%. In Weighted F-Measure, ResNet-101 showed a 2-3% performance improvement over ResNet-50, particularly on the CAMO dataset. The early stopping technique effectively reduced overfitting without sacrificing accuracy, especially for ResNet-101.

Item Type:	Thesis (S1)
Uncontrolled Keywords:	Deteksi Objek, Deteksi Objek Berkamuflase, Model CNN, ResNet-50, ResNet-101, Object Detection, Camouflage Object Detection, CNN Model, ResNet-50, ResNet-101.
Subjects:	L Education > L Education (General) L Education > LB Theory and practice of education L Education > LB Theory and practice of education > LB1501 Primary Education
Divisions:	UPI Kampus cibiru > S1 Rekayasa Perangkaat Lunak
Depositing User:	Dimas Aditya Permana
Date Deposited:	30 Aug 2024 02:57
Last Modified:	30 Aug 2024 02:57
URI:	http://repository.upi.edu/id/eprint/121501

Actions (login required)

View Item