Advanced

Demosaising using a Convolutional Neural Network approach

Dammer, Karin LU and Grosz, Ronja (2017) In Master's Theses in Mathematical Sciences FMA820 20171
Mathematics (Faculty of Engineering)
Abstract
This thesis is about investigating the feasibility to use convolutional neural networks as a demosaicing approach. Three loss methods and different layer structures have been evaluated as well as changing different parameters and layers in the convolutional neural network to find which changes are beneficial to make a neural network perform demosaicing well.

The convolutional neural network has been compared to a fully convolutional neural network, a multilayer perceptron and the Hamilton Adams demosaicing algorithm. The prospect of demosaicing raw image sensor data and images with noise was also investigated.

The conclusion is that a convolutional neural network can indeed perform demosaicing with good results, even when using a... (More)
This thesis is about investigating the feasibility to use convolutional neural networks as a demosaicing approach. Three loss methods and different layer structures have been evaluated as well as changing different parameters and layers in the convolutional neural network to find which changes are beneficial to make a neural network perform demosaicing well.

The convolutional neural network has been compared to a fully convolutional neural network, a multilayer perceptron and the Hamilton Adams demosaicing algorithm. The prospect of demosaicing raw image sensor data and images with noise was also investigated.

The conclusion is that a convolutional neural network can indeed perform demosaicing with good results, even when using a small and less complex network. The convolutional neural network was also able to demosaic raw images as well as remove noise from images, although with not as good result as when demosaicing artificial data. The convolutional neural network on average performed demosaicing with a peak signal to noise ratio of 34 dB. This compares to the Hamilton Adams method that has a peak signal to noise ratio of 37 dB, although when measured with the structural similarity our method performs better than the Hamilton Adams method. (Less)
Please use this url to cite or link to this publication:
author
Dammer, Karin LU and Grosz, Ronja
supervisor
organization
course
FMA820 20171
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Convolutional neural networks, fully convolutional neural network, demosaicing, image sensor data, noise reduction
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMA-3321-2017
ISSN
1404-6342
other publication id
2017:E30
language
English
id
8916452
date added to LUP
2017-06-22 16:17:25
date last changed
2017-06-22 16:17:25
@misc{8916452,
  abstract     = {This thesis is about investigating the feasibility to use convolutional neural networks as a demosaicing approach. Three loss methods and different layer structures have been evaluated as well as changing different parameters and layers in the convolutional neural network to find which changes are beneficial to make a neural network perform demosaicing well.
 
The convolutional neural network has been compared to a fully convolutional neural network, a multilayer perceptron and the Hamilton Adams demosaicing algorithm. The prospect of demosaicing raw image sensor data and images with noise was also investigated.
 
The conclusion is that a convolutional neural network can indeed perform demosaicing with good results, even when using a small and less complex network. The convolutional neural network was also able to demosaic raw images as well as remove noise from images, although with not as good result as when demosaicing artificial data. The convolutional neural network on average performed demosaicing with a peak signal to noise ratio of 34 dB. This compares to the Hamilton Adams method that has a peak signal to noise ratio of 37 dB, although when measured with the structural similarity our method performs better than the Hamilton Adams method.},
  author       = {Dammer, Karin and Grosz, Ronja},
  issn         = {1404-6342},
  keyword      = {Convolutional neural networks,fully convolutional neural network,demosaicing,image sensor data,noise reduction},
  language     = {eng},
  note         = {Student Paper},
  series       = {Master's Theses in Mathematical Sciences},
  title        = {Demosaising using a Convolutional Neural Network approach},
  year         = {2017},
}