Salient pixels and dimensionality reduction for display of multi/hyperspectral images

  • Authors:

  • Steven Le Moan;Ferdinand Deger;Alamin Mansouri;Yvon Voisin;Jon Y. Hardeberg

  • Affiliations:

  • Laboratoire d'Electronique, Informatique et Image, Université de Bourgogne, Auxerre, France, The Norwegian Color Research Laboratory, Gjøvik University College, Norway;Laboratoire d'Electronique, Informatique et Image, Université de Bourgogne, Auxerre, France, The Norwegian Color Research Laboratory, Gjøvik University College, Norway;Laboratoire d'Electronique, Informatique et Image, Université de Bourgogne, Auxerre, France;Laboratoire d'Electronique, Informatique et Image, Université de Bourgogne, Auxerre, France;The Norwegian Color Research Laboratory, Gjøvik University College, Norway

  • Venue:
  • ICISP'12 Proceedings of the 5th international conference on Image and Signal Processing
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

Dimensionality Reduction (DR) of spectral images is a common approach to different purposes such as visualization, noise removal or compression. Most methods such as PCA or band selection use either the entire population of pixels or a uniformly sampled subset in order to compute a projection matrix. By doing so, spatial information is not accurately handled and all the objects contained in the scene are given the same emphasis. Nonetheless, it is possible to focus the DR on the separation of specific Objects of Interest (OoI), simply by neglecting all the others. In PCA for instance, instead of using the variance of the scene in each spectral channel, we show that it is more efficient to consider the variance of a small group of pixels representing several OoI, which must be separated by the projection. We propose an efficient method based on saliency to automatically identify OoI and extract only a few relevant pixels to enhance the separation foreground/background in the DR process.