PENGEMBANGAN MODEL THERMAL RESIDUALNET-128 UNTUK DETEKSI OBJEK DALAM CITRA TERMAL MENGGUNAKAN PEMBELAJARAN RESIDUAL

Miftahul Falah, - and Galura Muhammad Suranegara, - and Ichwan Nul Ichsan, - (2025) PENGEMBANGAN MODEL THERMAL RESIDUALNET-128 UNTUK DETEKSI OBJEK DALAM CITRA TERMAL MENGGUNAKAN PEMBELAJARAN RESIDUAL. S1 thesis, Universitas Pendidikan Indonesia.

Abstract

Deteksi objek untuk kendaraan otonom menghadapi tantangan dalam kondisi pencahayaan rendah. Citra termal menawarkan solusi yang efektif, namun arsitektur Convolutional Neural Network (CNN) konvensional sering mengalami penurunan performa pada jaringan dalam. Penelitian ini bertujuan mengembangkan dan mengevaluasi model Thermal ResidualNet-128 (TRNet-128), arsitektur CNN dengan mekanisme pembelajaran residual untuk mendeteksi objek manusia dan pengendara sepeda motor. Dataset yang digunakan terdiri dari 1.008 gambar citra termal grayscale. Performa model TRNet-128 dievaluasi menggunakan metrik accuracy, precision, recall, F1-score, confusion matrix, dan mean Average Precision (mAP), serta dibandingkan dengan model CNN sederhana tanpa residual block. Hasil menunjukkan model TRNet-128, setelah augmentasi data untuk mengatasi overfitting, mencapai akurasi validasi 99,86% dan mAP 0,99977, jauh lebih tinggi dibandingkan model CNN sederhana yang hanya 0,99652. Kesimpulannya, integrasi residual blocks pada TRNet-128 secara efektif meningkatkan akurasi klasifikasi dan ketepatan deteksi objek dalam citra termal. ----- Object detection for autonomous vehicles faces challenges in low-light conditions. Thermal imagery offers an effective solution; however, conventional Convolutional Neural Network (CNN) architectures often experience performance degradation in deep networks. This study aims to develop and evaluate the Thermal ResidualNet-128 (TRNet-128) model, a CNN architecture equipped with a residual learning mechanism to detect human and motorcycle rider objects. The dataset used consists of 1,008 grayscale thermal images. The performance of the TRNet-128 model is evaluated using metrics such as accuracy, precision, recall, F1-score, confusion matrix, and mean Average Precision (mAP), and compared to a simple CNN model without residual block. Results show that the TRNet-128 model, after data augmentation to mitigate overfitting, achieved a validation accuracy of 99.86% and an mAP of 0.99977, significantly outperforming the simple CNN model, which only reached 0.99652. In conclusion, the integration of residual blocks in the TRNet-128 architecture effectively enhances classification accuracy and object detection precision in thermal images.

	Text S_SISTEL_2104755_Title.pdf Download (293kB)
	Text S_SISTEL_2104755_Chapter1.pdf Download (243kB)
	Text S_SISTEL_2104755_Chapter2.pdf Restricted to Staf Perpustakaan Download (267kB) | Request a copy
	Text S_SISTEL_2104755_Chapter3.pdf Download (388kB)
	Text S_SISTEL_2104755_Chapter4.pdf Restricted to Staf Perpustakaan Download (343kB) | Request a copy
	Text S_SISTEL_2104755_Chapter5.pdf Download (227kB)
	Text S_SISTEL_2104755_Appendix.pdf Restricted to Staf Perpustakaan Download (442kB) | Request a copy

Item Type:	Thesis (S1)
Additional Information:	https://scholar.google.com/citations?user=NsVZCLoAAAAJ&hl=en&authuser=3 ID SINTA Dosen Pembimbing Galura Muhammad Suranegara: 6703764 Ichwan Nul Ichsan: 6721201
Uncontrolled Keywords:	Deteksi Objek, Citra Termal, Convolutional Neural Network, Pembelajaran Residual, TRNet-128 Object Detection, Thermal Imagery, Convolutional Neural Network, Residual Learning, TRNet-128
Subjects:	T Technology > T Technology (General)
Divisions:	UPI Kampus Purwakarta > S1 Sistem Telekomunikasi
Depositing User:	Miftahul Falah
Date Deposited:	10 Sep 2025 08:02
Last Modified:	10 Sep 2025 08:02
URI:	http://repository.upi.edu/id/eprint/135573

Actions (login required)

View Item