A study of video denoising based on deep learning
Video denoising is a fundamental low-level vision task which means videos corrupted by noise are restored and high-quality videos are obtained. With deep learning techniques applied in the computer vision area, models based on deep neural network began to replace traditional algorithms. In this d...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/180018 |
| Summary: | Video denoising is a fundamental low-level vision task which means videos corrupted
by noise are restored and high-quality videos are obtained. With deep
learning techniques applied in the computer vision area, models based on deep
neural network began to replace traditional algorithms. In this dissertation, how
video denoising techniques developed in the past few years and other related
knowledge are summarized after browsing a large amount of literature. Then
two supervised methods (DVDnet and PaCNet) and two unsupervised methods
(RFR and UDVD) are explained and illustrated. To evaluate their ability to remove
AWGN, benchmarks DVAIS and Set8 are used. Then DVDnet which is
the best considering both denoised results and running time is chosen for further
experiments. A self-collected low-light dataset NIGHT and a realistic noise
dataset CRVD are used to test the generalization ability of DVDnet. Quantitative
and visual results indicate that DVDnet has remarkable denoising performance
and acceptable generalization ability. |
|---|