Classification of Color Pigments in Hyperspectral Images (bibtex)
by Christian Asinger
Abstract:
Image spectroscopy in ultraviolet, visible and infrared regions provides a non-invasive method for analyzing paintings. It is used by restorers and art-historians to get valuable information about works of art without causing any damage to them. The purpose of this report is to analyze 15 different color pigments and three types of binder by means of near infrared hyperspectral images taken from 15 testpanels. After calibration three feature reduction methods, principal component analysis, block-based principal component analysis and linear discriminant analysis, are applied to the near-IR spectroscopic imaging data of the testpanels to reduce the originally 180 features to improve classification performance. Both, principal component analysis and linear discriminant analysis, are linear transformations (in contrast to e.g. Kernel PCA) that map the original data into a new vector space where certain constraints are met. On the basis of the outcomes of the feature reduction step k-nearest neighbor classification with leave-one-out cross validation is used to accomplish three different classification tasks. These three tasks contain the identification of the three binders, the identification of the color pigments for each binder separately and the identification of color pigments and binders together. The results of our work reveal that it is possible to distinguish the three types of binder as well as the color pigments for each binder separately with an accuracy of more than 90 percent. Classifying the 15 color pigments along with the three binders however results in a lower accuracy of at most 45 percent.
Reference:
Classification of Color Pigments in Hyperspectral Images (Christian Asinger), Technical report, PRIP, TU Wien, 2004.
Bibtex Entry:
@TechReport{TR093,
  author =	 "Christian Asinger",
  title =	 "Classification of Color Pigments in Hyperspectral
                  Images",
  institution =	 "PRIP, TU Wien",
  number =	 "PRIP-TR-093",
  year =	 "2004",
  url =		 "ftp://ftp.prip.tuwien.ac.at/pub/publications/trs/tr93.pdf",
  abstract =	 "Image spectroscopy in ultraviolet, visible and
                  infrared regions provides a non-invasive method for
                  analyzing paintings. It is used by restorers and
                  art-historians to get valuable information about
                  works of art without causing any damage to them. The
                  purpose of this report is to analyze 15 different
                  color pigments and three types of binder by means of
                  near infrared hyperspectral images taken from 15
                  testpanels. After calibration three feature
                  reduction methods, principal component analysis,
                  block-based principal component analysis and linear
                  discriminant analysis, are applied to the near-IR
                  spectroscopic imaging data of the testpanels to
                  reduce the originally 180 features to improve
                  classification performance. Both, principal
                  component analysis and linear discriminant analysis,
                  are linear transformations (in contrast to
                  e.g. Kernel PCA) that map the original data into a
                  new vector space where certain constraints are
                  met. On the basis of the outcomes of the feature
                  reduction step k-nearest neighbor classification
                  with leave-one-out cross validation is used to
                  accomplish three different classification
                  tasks. These three tasks contain the identification
                  of the three binders, the identification of the
                  color pigments for each binder separately and the
                  identification of color pigments and binders
                  together. The results of our work reveal that it is
                  possible to distinguish the three types of binder as
                  well as the color pigments for each binder
                  separately with an accuracy of more than 90
                  percent. Classifying the 15 color pigments along
                  with the three binders however results in a lower
                  accuracy of at most 45 percent.",
}
Powered by bibtexbrowser