technique for searching images in a spectral image database
DESCRIPTION
Optics in Engineering, OIE’03, Saariselkä, FINLAND, August 7, 2003. Technique for Searching Images in a Spectral Image Database. Markku Hauta-Kasari, Kanae Miyazawa *, Jussi Parkkinen, and Timo Jaaskelainen University of Joensuu Color Group - PowerPoint PPT PresentationTRANSCRIPT
Technique for Searching Images in a Spectral Image Database
Markku Hauta-Kasari, Kanae Miyazawa *, Jussi Parkkinen, and Timo Jaaskelainen
University of Joensuu Color GroupDepartment of Computer Science, Department of Physics
University of Joensuu, FINLAND
*Department of Information and Computer Sciences, Toyohashi University of Technology, JAPAN
Kanae Miyazawa: PhD from Prof. Toyooka lab., 2001-2003 in JoensuuMarkku Hauta-Kasari: Visiting researcher 1996-1998 at Prof. Toyooka lab.
Optics in Engineering, OIE’03, Saariselkä, FINLAND, August 7, 2003
Contents
1. Introduction2. Searching Technique3. Experimental Results4. Conclusions
IntroductionAccurate color representation:
spectral color representation
Spectral image: high-accurate color image, large amount of data
Application fields, for example:quality control, telemedicine, e-commerce, electronic museums
Spectral image databases expanding: fast methods for searching images will be needed in the future
The Aim of this StudyTo propose a technique for searching images in a spectral image database
Experimental data76 real-world spectral images
• usually, color filters with fixed transmittances are used:– filters must be physically changed to the system – the transmittance of the filter cannot be changed
• computational color filter design:– the transmittance of the filter can be adaptive to an application
spectral imaging systems that can use arbitrary rewritable color filters are needed
IntroductionSpectral image databases expanding: fast methods for searching images will be needed in
the future
The Aim of this StudyTo propose a technique for searching images in a spectral image database
Experimental data76 real-world spectral images
Searching Technique
Histogram database creation
1. Select spectra randomly, train a SOM2. Calculate BMU-images and BMU-histograms
Search
3. Select a wanted image, calculate BMU-image and BMU-histogram4. Calculate histogram similarities5. Order the search results based on the similarities
Filter = Light Source
Sample
Self-Organized Map
Filter = Light Source
Sample
SOM-units in CIELAB
Filter = Light Source
Sample
Example of BMU-image
Filter = Light Source
Sample
Example of BMU-histogram
Filter = Light Source
Sample
Example of Search
Filter = Light Source
Sample
Histogram Differences
Conclusions• The histogram database is generated once for a certain spectral
image database. This is a time consuming phase.
• The searching technique is fast and it preserves the spectral color information. The searching time for synthetically generated 1000 spectral image test set was 1 second (Matlab, Linux PC)
Future: • to test more features for BMU-histogram similarity calculation• to add textural features to the technique