Leveraging IIR State in Deep Neural Networks for Video Denoising
(2025) In Master's Theses in Mathematical Sciences FMAM05 20251Mathematics (Faculty of Engineering)
- Abstract
- In recent years, machine learning models have changed the field of video denoising. Most state-of-the-art models utilize multiple frames as input to denoise a single frame, which limits the temporal context window and often requires more computational power. Another, less explored, approach is to use recurrence to gain a hidden memory state in which temporal information can be stored. While the former approach scales poorly with larger temporal context windows, the latter lacks proper analysis of its relevance and trade-offs. In this master’s thesis, we establish experiments to better understand the feasibility of using recurrence for video denoising, and to study the artifacts that this approach can introduce. Furthermore, we attempt to... (More)
- In recent years, machine learning models have changed the field of video denoising. Most state-of-the-art models utilize multiple frames as input to denoise a single frame, which limits the temporal context window and often requires more computational power. Another, less explored, approach is to use recurrence to gain a hidden memory state in which temporal information can be stored. While the former approach scales poorly with larger temporal context windows, the latter lacks proper analysis of its relevance and trade-offs. In this master’s thesis, we establish experiments to better understand the feasibility of using recurrence for video denoising, and to study the artifacts that this approach can introduce. Furthermore, we attempt to construct objective metrics to measure the flickering, ghosting, and trailing, that temporal filters can give rise to. We used a dataset captured by statically mounted surveillance cameras, upon which we added synthetic noise using a realistic noise model. Using the constructed metrics and standard performance metrics, we present differences in the use of these models. We also compare these models against a state-of-the-art traditional video denoising model, and a machine learning model that solely performs spatial noise filtering. The results show that the recurrent model outperforms the other models in denoising the video, while preserving details and temporal consistency in static regions. This comes at the expense of worse performance in objects that move across the scene. The proposed metrics also show promise in objectively capturing effects that standard metrics miss. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9192779
- author
- Sjölin, Max LU and Bergrahm, Erik LU
- supervisor
-
- Karl Åström LU
- Gustav Hanning LU
- organization
- course
- FMAM05 20251
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- video denoising, IIR memory state, temporal consistency, video quality assessment, recurrent neural networks, ghosting, trailing, flickering
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMA-3578-2025
- ISSN
- 1404-6342
- other publication id
- 2025:E28
- language
- English
- id
- 9192779
- date added to LUP
- 2025-09-15 11:14:07
- date last changed
- 2025-09-15 11:14:07
@misc{9192779,
abstract = {{In recent years, machine learning models have changed the field of video denoising. Most state-of-the-art models utilize multiple frames as input to denoise a single frame, which limits the temporal context window and often requires more computational power. Another, less explored, approach is to use recurrence to gain a hidden memory state in which temporal information can be stored. While the former approach scales poorly with larger temporal context windows, the latter lacks proper analysis of its relevance and trade-offs. In this master’s thesis, we establish experiments to better understand the feasibility of using recurrence for video denoising, and to study the artifacts that this approach can introduce. Furthermore, we attempt to construct objective metrics to measure the flickering, ghosting, and trailing, that temporal filters can give rise to. We used a dataset captured by statically mounted surveillance cameras, upon which we added synthetic noise using a realistic noise model. Using the constructed metrics and standard performance metrics, we present differences in the use of these models. We also compare these models against a state-of-the-art traditional video denoising model, and a machine learning model that solely performs spatial noise filtering. The results show that the recurrent model outperforms the other models in denoising the video, while preserving details and temporal consistency in static regions. This comes at the expense of worse performance in objects that move across the scene. The proposed metrics also show promise in objectively capturing effects that standard metrics miss.}},
author = {{Sjölin, Max and Bergrahm, Erik}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Leveraging IIR State in Deep Neural Networks for Video Denoising}},
year = {{2025}},
}