A Novel Perceptual Metric in Deep Learning
(2020) In Master's Theses in Mathematical Sciences FMAM05 20201Mathematics (Faculty of Engineering)
- Abstract
- Loss functions are a crucial part of image processing when using modern neural networks, trained with stochastic gradient descent. There exist multiple loss functions today, some claiming to be perceptual. Researchers at NVIDIA recently published a proposal of such a metric called FLIP. This led to our work on this thesis where we present a comparison between multiple loss functions, both established ones but also a completely new one. The loss functions were subjected to two problems, denoising RGB images and reconstructing MR images.
The central question is how does one evaluate the evaluation? If we train a network with the same architecture but with different loss functions, and then measure the performance with one of these loss... (More) - Loss functions are a crucial part of image processing when using modern neural networks, trained with stochastic gradient descent. There exist multiple loss functions today, some claiming to be perceptual. Researchers at NVIDIA recently published a proposal of such a metric called FLIP. This led to our work on this thesis where we present a comparison between multiple loss functions, both established ones but also a completely new one. The loss functions were subjected to two problems, denoising RGB images and reconstructing MR images.
The central question is how does one evaluate the evaluation? If we train a network with the same architecture but with different loss functions, and then measure the performance with one of these loss functions, the result would most likely be biased. We therefore present a comparison between loss functions where we both visually and numerically evaluate the results as well as conducting user studies. We find that the weighted version of LPIPS + l2 and MS-SSIM are especially good loss functions for mentioned problems, and the newly proposed metric FLIP performed well. (Less) - Popular Abstract (Swedish)
- Förmågan att återskapa förvrängda bilder på ett effektivt sätt har länge varit ett mål inom datorgrafik och bildanalys. I vårt examensarbete, i samarbete med NVIDIA, har vi tittat på ett nytt tillvägagångsätt för detta problem inom maskininlärning och jämfört detta med redan existerande metoder.
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9022150
- author
- Engman, Johanna LU and Nilsson, Hanna LU
- supervisor
- organization
- alternative title
- A comparison of Loss Functions for Image Denoising and Reconstruction
- course
- FMAM05 20201
- year
- 2020
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Loss function, Metric, Deep learning, Denoising, MRI reconstruction, U-net, Machine learning, Image Analysis
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMA-3413-2020
- ISSN
- 1404-6342
- other publication id
- 2020:E33
- language
- English
- id
- 9022150
- date added to LUP
- 2020-06-29 14:40:50
- date last changed
- 2020-06-29 14:40:50
@misc{9022150, abstract = {{Loss functions are a crucial part of image processing when using modern neural networks, trained with stochastic gradient descent. There exist multiple loss functions today, some claiming to be perceptual. Researchers at NVIDIA recently published a proposal of such a metric called FLIP. This led to our work on this thesis where we present a comparison between multiple loss functions, both established ones but also a completely new one. The loss functions were subjected to two problems, denoising RGB images and reconstructing MR images. The central question is how does one evaluate the evaluation? If we train a network with the same architecture but with different loss functions, and then measure the performance with one of these loss functions, the result would most likely be biased. We therefore present a comparison between loss functions where we both visually and numerically evaluate the results as well as conducting user studies. We find that the weighted version of LPIPS + l2 and MS-SSIM are especially good loss functions for mentioned problems, and the newly proposed metric FLIP performed well.}}, author = {{Engman, Johanna and Nilsson, Hanna}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master's Theses in Mathematical Sciences}}, title = {{A Novel Perceptual Metric in Deep Learning}}, year = {{2020}}, }