Neural Style Transfer Pada Video Secara Realtime Menggunakan Berbagai Model

Ramandanta, Achmad Khosyi' Assajjad (2025) Neural Style Transfer Pada Video Secara Realtime Menggunakan Berbagai Model. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025211007-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Neural Style Transfer (NST) memungkinkan penerapan gaya artistik dari sebuah gambar ke gambar atau video lainnya. Penelitian ini menganalisis kecepatan komputasi dan kualitas visual NST pada video menggunakan tiga model arsitektur yaitu VGG19, CNN Feed-forward, dan model dari TensorFlow Hub. Pengujian dilakukan pada berbagai resolusi video (1080p, 720p, 480p) dan jenis Graphics Processing Unit (GPU) (T4, L4, A100). Hasil menunjukkan bahwa model CNN Feed-forward menawarkan performa tercepat, diikuti oleh model TensorFlow Hub, sementara VGG19 adalah yang paling lambat. Peningkatan resolusi video secara konsisten mengurangi throughput dan meningkatkan rasio realtime pada semua konfigurasi. Konfigurasi paling mendekati realtime adalah resolusi 480p dengan GPU A100, mencapai rasio realtime 1,066 untuk CNN Feed-forward dan 1,445 untuk TensorFlow Hub. Analisis kualitas visual menunjukkan trade-off signifikan antar model, dimana CNN Feed-forward memiliki rasio content loss:style loss 99%:1%, VGG19 5%:95%, dan TensorFlow Hub 1%:99%. Evaluasi subjektif dari 31 responden menunjukkan korelasi yang konsisten dengan metrik objektif, dimana 74,2% responden memilih CNN Feed-forward sebagai hasil terbaik karena kemampuannya mempertahankan kejelasan objek dan konsistensi temporal yang baik. Penelitian ini memberikan panduan dalam memilih model dan konfigurasi yang optimal untuk aplikasi NST video secara realtime berdasarkan aspek performa komputasi dan kualitas visual.
===============================================================================================================================
Neural Style Transfer (NST) enables the application of artistic styles from one image to another image or video. This research analyzes computational speed and visual quality of NST on video using three model architectures: VGG19, CNN Feed-forward, and TensorFlow Hub model. Testing was conducted on various video resolutions (1080p, 720p, 480p) and Graphics Processing Unit (GPU) types (T4, L4, A100). Results show that CNN Feed-forward model offers the fastest performance, followed by TensorFlow Hub model, while VGG19 is the slowest. Increasing video resolution consistently reduces throughput and increases realtime ratio across all configurations. The configuration closest to realtime is 480p resolution with A100 GPU, achieving realtime ratios of 1.066 for CNN Feed-forward and 1.445 for TensorFlow Hub. Visual quality analysis reveals significant trade-offs between models, where CNN Feed-forward has content loss:style loss ratio of 99%:1%, VGG19 5%:95%, and TensorFlow Hub 1%:99%. Subjective evaluation from 31 respondents shows consistent correlation with objective metrics, where 74.2% of respondents chose CNN Feed-forward as the best result due to its ability to maintain object clarity and good temporal consistency. This research provides guidance in selecting optimal models and configurations for realtime video NST applications based on computational performance and visual quality aspects.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	CNN Feed-forward, Neural Style Transfer, TensorFlow Hub, VGG19, Video, CNN Feed-forward, Neural Style Transfer, TensorFlow Hub, VGG19, Video
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA76.9.I52 Information visualization T Technology > T Technology (General) > T385 Visualization--Technique
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Achmad Khosyi' Assajjad Ramandanta
Date Deposited:	29 Jul 2025 10:29
Last Modified:	29 Jul 2025 10:29
URI:	http://repository.its.ac.id/id/eprint/122723

Actions (login required)

View Item