Pengaruh Teknik Augmentasi Citra Wajah Terhadap Performa Deteksi Wajah YOLO Versi 8 Pada Video 360 Derajat

Ardy, Rizky Damara (2025) Pengaruh Teknik Augmentasi Citra Wajah Terhadap Performa Deteksi Wajah YOLO Versi 8 Pada Video 360 Derajat. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025231004-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 April 2027. Download (110MB) | Request a copy

Abstract

Teknologi video 360 derajat telah berkembang pesat dan diintegrasikan ke dalam berbagai aplikasi seperti hiburan, pendidikan, dan keamanan, di mana deteksi wajah yang akurat dan efisien menjadi sangat krusial untuk mendukung navigasi, interaksi, dan analisis. Namun, deteksi wajah dalam video 360 menghadapi tantangan unik dibandingkan video konvensional, terutama karena distorsi visual akibat proyeksi equirectangular yang membuat wajah tampak tidak proporsional dan sulit dikenali. Penelitian ini menyelidiki efektivitas teknik augmentasi citra dalam meningkatkan performa deteksi wajah menggunakan algoritma YOLOv8 pada video 360, dengan memanfaatkan dataset WIDERFace dan berbagai teknik augmentasi seperti fotometri (HSV, derau, kecerahan, BGR), geometri (rotasi, translasi, skala, shear, prespektif, flip horisontal vertikal, equirectangular) blending (mozaik, mixup, copy-paste) dan pipeline. Penelitian ini memberikan kontribusi penting dalam memahami pengaruh teknik augmentasi citra terhadap performa deteksi wajah dan menawarkan tolak ukur untuk implementasi deteksi wajah yang lebih efektif dalam video 360 derajat. Hasil penelitian menunjukkan bahwa teknik augmentasi blending meningkatkan performa deteksi wajah pada video 360 derajat. Rangkaian augmentasi ini memberikan peningkatan yang lebih besar dibandingkan dengan metode augmentasi fotometri, geometri, dan pipeline. Secara khusus, teknik blending meningkatkan performa deteksi wajah sebesar 3.691% pada Cube Map Projection (CMP) dan 6.164% pada Equirectangular Projection (ERP). Selain itu, teknik augmentasi fotometri menunjukkan peningkatan sebesar 2.632% pada CMP, sementara teknik pipeline mencatat peningkatan sebesar 1.346% pada ERP, jika dibandingkan dengan augmentasi standar dari YOLOv8. Sementara itu, teknik augmentasi yang paling banyak mendeteksi wajah adalah teknik pipeline, yang menunjukkan peningkatan sebesar 3.04% pada CMP dan 9.70% pada ERP dibandingkan dengan augmentasi standar.
===================================================================================================================================
360-degree video technology has rapidly evolved and has been integrated into various applications such as entertainment, education, and security, where accurate and efficient face detection is crucial to support navigation, interaction, and analysis. However, face detection in 360-degree video faces unique challenges compared to conventional video, particularly due to visual distortions caused by the equirectangular projection that make faces appear disproportionate and difficult to recognize. This research investigates the effectiveness of image augmentation techniques in enhancing face detection performance using the YOLOv8 algorithm on 360-degree video, leveraging the WIDERFace dataset and various augmentation techniques such as photometric (HSV, noise, brightness, BGR), geometric (rotation, translate, scale, shear, perspective, flip horizontal vertical, equirectangular) blending (mosaic, mix-up, copy-paste), and pipeline. This study makes significant contributions to understanding the impact of image augmentation techniques on face detection performance and offers benchmarks for more effective face detection implementation in 360-degree video. The results indicate that the blending augmentation technique improves face detection performance in 360-degree videos. This series of augmentations provides a greater enhancement compared to photometric, geometric, and pipeline augmentation methods. Specifically, the blending technique increases face detection performance by 3.691% in Cube Map Projection (CMP) and 6.164% in Equirectangular Projection (ERP). Additionally, the photometric augmentation technique shows an increase of 2.632% in CMP, while the pipeline technique records an increase of 1.346% in ERP compared to the standard augmentation from YOLOv8. Meanwhile, the augmentation technique that detects faces the most is the pipeline technique, which shows an increase of 3.04% on CMP and 9.70% on ERP compared to standard augmentation.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Augmentasi citra, Dataset WiderFace, Deteksi wajah, Distorsi visual, Proyeksi equirectangular, Video 360 derajat, YOLOv8; Equirectangular projection, Face detection, Image augmentation,Visual distortion, WiderFace dataset, YOLOv8, 360-degree video
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1650 Face recognition. Optical pattern recognition.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Rizky Damara Ardy
Date Deposited:	03 Feb 2025 03:32
Last Modified:	03 Feb 2025 03:32
URI:	http://repository.its.ac.id/id/eprint/117751

Actions (login required)

View Item