Advanced

Classification of Textures Using Convolutional Neural Networks

Persson, Ivar LU (2017) In Master's Theses in Mathematical Sciences FMA820 20171
Mathematics (Faculty of Engineering)
Abstract
This Master's thesis has concerned the segmentation and classification of background textures in images. In order to segment the images we have used the SLIC algorithm to create superpixels. These are a sort of over segmentation of the image where pixels close to each other and similar in colour are considered to be the same texture. All superpixels were then classified using a convolutional neural network which was trained as a part of this thesis. As this network had about 30% errors a second stage was added to the classification program, a spatial bias. The first attempt at this spatial bias used the neighbouring superpixels' classification in order to make the image more homogeneous. Secondly, as a comparison, a neural network was also... (More)
This Master's thesis has concerned the segmentation and classification of background textures in images. In order to segment the images we have used the SLIC algorithm to create superpixels. These are a sort of over segmentation of the image where pixels close to each other and similar in colour are considered to be the same texture. All superpixels were then classified using a convolutional neural network which was trained as a part of this thesis. As this network had about 30% errors a second stage was added to the classification program, a spatial bias. The first attempt at this spatial bias used the neighbouring superpixels' classification in order to make the image more homogeneous. Secondly, as a comparison, a neural network was also trained as the spatial bias.

When using the neural network followed by the neural network for spatial biases the errors decreased to a little more than 10%, while the ordinary spatial bias only decreased error by a few percent. Even though the best performing network had a low error rate we were not able to replicate these results with unknown images as the network most likely severely overfitted to the relatively small training set. The method in it self showed potential and with more training data and possibly smaller superpixels we could get more consistent results between training images and other unknown data. (Less)
Please use this url to cite or link to this publication:
author
Persson, Ivar LU
supervisor
organization
course
FMA820 20171
year
type
H2 - Master's Degree (Two Years)
subject
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMA-3325-2017
ISSN
1404-6342
other publication id
2017:E40
language
English
id
8917826
date added to LUP
2017-06-22 16:06:16
date last changed
2017-06-22 16:06:16
@misc{8917826,
  abstract     = {This Master's thesis has concerned the segmentation and classification of background textures in images. In order to segment the images we have used the SLIC algorithm to create superpixels. These are a sort of over segmentation of the image where pixels close to each other and similar in colour are considered to be the same texture. All superpixels were then classified using a convolutional neural network which was trained as a part of this thesis. As this network had about 30% errors a second stage was added to the classification program, a spatial bias. The first attempt at this spatial bias used the neighbouring superpixels' classification in order to make the image more homogeneous. Secondly, as a comparison, a neural network was also trained as the spatial bias.
 
 When using the neural network followed by the neural network for spatial biases the errors decreased to a little more than 10%, while the ordinary spatial bias only decreased error by a few percent. Even though the best performing network had a low error rate we were not able to replicate these results with unknown images as the network most likely severely overfitted to the relatively small training set. The method in it self showed potential and with more training data and possibly smaller superpixels we could get more consistent results between training images and other unknown data.},
  author       = {Persson, Ivar},
  issn         = {1404-6342},
  language     = {eng},
  note         = {Student Paper},
  series       = {Master's Theses in Mathematical Sciences},
  title        = {Classification of Textures Using Convolutional Neural Networks},
  year         = {2017},
}