Video Compositing Menggunakan Spatio Temporal Poisson Blending Berbasis Temporal Alpha Matte

Koeshardianto, Meidya (2020) Video Compositing Menggunakan Spatio Temporal Poisson Blending Berbasis Temporal Alpha Matte. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

	Text 07111460010008-Disertation.pdf Restricted to Repository staff only Download (42MB) | Request a copy
	Text 07111460010008-Disertation.pdf Restricted to Repository staff only Download (21MB) | Request a copy

Abstract

Teknik ekstraksi obyek foreground dari citra background dibutuhkan dalam berbagai aplikasi movie maker khususnya dibidang animasi dan visual efek. Salah satu teknik ekstraksi yang sering digunakan adalah image segmentation. Namun hasilnya tidak dapat memisahkan obyek foreground pada citra secara detail. Image matting menjadi solusi untuk mengatasi permasalahan akurasi ekstraksi obyek foreground pada citra. Pada proses matting diperlukan constraint (scribbles atau trimap) sebagai label pada citra yang menunjukkan foreground dan background. Pada penelitian - penelitian sebelumnya, constraint diselesaikan secara manual sehingga proses labelling data video dilakukan per frame membutuhkan waktu dan tenaga yang kurang efisien. Pada penelitian Disertasi ini dibangun beberapa model constraint secara otomatis. Model tersebut antara lain menggunakan citra template (template matting), menggunakan fitur edge dan yang terakhir model dibangun menggunakan fitur DCT. Implementasi model matting otomatis menghasilkan error alpha matte sebesar 0.0336 menggunakan algoritma MAE apabila dibandingkan dengan metode manual. Error tersebut berdampak pada besarnya gradasi saat alpha matte diimplementasikan pada video compositing. Gradasi hasil matting otomatis menjadi fokus penelitian selanjutnya karena proses menghasilkan alpha matte sebelumnya dilakukan per frame. Untuk memperbaiki alpha matte pada proses video compositing dibangun model spatio temporal poisson blending yang implementasinya melibatkan alpha matte pada frame sebelum dan sesudahnya (temporal alpha matte).
===============================================================================================================================
Many applications in movie maker, especially in visual effect animation, have difficulties in putting the object on the real background scene. Extracting foreground technique has been used to obtain the object. Therefore, it could be composited with the background and taken in the real condition. The separation process between the foreground and background can be done using some techniques, such as image segmentation. Nevertheless, image egmentation results still cannot separate objects accurately. Image matting has been proposed as a solution to solve this problem of object separation. The matting methods need defined foreground and background constraints before applying matting. These methods have advantages that foreground and background can be determined accurately, but it requires effort for manual labeling using trimap or scribbles, especially on video dataset. In this research implies a supervised approach without manual guidance. Scribbles will be created automatically using in several ways, starting from forming scribbles using template images or called the template matting. Further research scribbles were built using the edge feature, and finally, the
scribbles model was built using the DCT feature. The implementation of the automatic matting model produces an alpha matte error of 0.0336 using the MAE algorithm when compared with the manual method. This error affects the amount of gradation when alpha matte is implemented in video com positing. Gradation of automatic matting results becomes the focus of further research because the process of producing alpha matte was previously done per frame. For improving alpha matte in the video compositing process, a spatia termporal poisson blending model was built, the implementation of which involves alpha matte in the before and after frames (temporal alpha matte).

Item Type:	Thesis (Doctoral)
Additional Information:	RDE 006.42 Koe v-1
Uncontrolled Keywords:	matting, compositing, template matting, temporal alpha matte, automatic matting, spatio temporal poisson blending
Subjects:	T Technology > TR Photography > TR267.733.M85 Multispectral imaging T Technology > TR Photography > TR845 Cinematography. Video recordings. T Technology > TR Photography > TR897.7 Computer animation
Divisions:	Faculty of Electrical Technology > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Meidya Koeshardianto
Date Deposited:	14 Mar 2025 01:41
Last Modified:	14 Mar 2025 01:41
URI:	http://repository.its.ac.id/id/eprint/75214

Actions (login required)

View Item