1. Rolando Raqueno, Advisor Credits for Winter Quarter, 2002: 2
Visual Enhancement of the Archimedes Palimpsest Using a Target Detection Algorithm or
Trying to read ancient Greek written over a thousand years ago that’s been burned, erased, overwritten, torn apart, put back together, and has mold growing on it.
By GaryHoffmann
Roger Easton, Advisor
Rolando Raqueno, Advisor
Credits for Winter Quarter, 2002: 2

2. Contents Introduction Specific Aims Background and Significance
Experimental Design and Methods
Resources and Environment
Timetable
Budget

3. Introduction Old manuscripts often damaged: fading of ink
presence of mold
deliberate defacement

4. Methods have been developed to restore
documents of cultural significance
3 band color image (left) and 6 band ‘Super-visual’ image (right)

5. Hyperspectral Imagery
Hundreds of bands
Provides more information than a multispectral image
Could be used to improve document restoration techniques
Target detection methods could be used on hyperspectral images of old manuscripts

6. Specific Aims
Continue Verification of Kyungsuk Lee’s algorithm for target detection
Test algorithm on different subsections of AVIRIS scene used during development
Different AVIRIS images
Different sensors - MISI, Hyperion, HYDICE, etc.
Possible thresholds that can be used to automate the algorithm

7. Specific Aims
Capture more hyperspectral imagery of Archimedes Palimpsest using tunable filter/CCD and an Analytical Spectral Device (ASD)
Sensor/illuminant geometry
Samples per line with ASD and lines for a given area
ASD spot size and time required to acquire data for a given area
Constant illumination
Transmission of glass plate
Compare to ink spectra previously measured
ASD dark image

8. Specific Aims Analyze data and implement algorithm on manuscript
Usable spectral features in ink
Use MODTRAN to find appearance of ink given all possible geometries
Apply algorithm to collected images

9. Background
The Archimedes Palimpsest

10. Method to locate underwriting:
Material identification with hyperspectral imagery
High spectral-dimensionality
Can differentiate materials that cannot be differentiated using sensors with fewer bands
Detection of materials made difficult by variations in illumination across image
Sensor/illuminant geometry
Atmospheric conditions (remote sensing)
Etc.

11. Glenn Healey’s Invariant Target Detection Method
Dimensionality of possible spectral radiance vectors for target less than dimensionality of sensor
Vector set is invariant to illumination (spans all possible appearances of target for different conditions)
Some set of basis vectors will span this set
Any target vector is linear combination of basis vectors
n = L(l) - Stiai

12. Determining invariant target space:
Use of physics based models
All parameters varied across all extremes
Apply to target reflectance vector
Resulting space invariant to illumination

13. Kyungsuk Lee’s Target Detection Method
Extends to subpixel scale
Find basis vectors for background, as well
Hypothesis test:
Each pixel is linear combination background basis vectors only (target not present)
Each pixel is linear combination of target basis vectors and background basis vectors (target present)
n0 = L(l) - Sbngn
n1 = L(l) - [Sbngn + Stmam]

14. Kyungsuk Lee’s Target Detection Method
Find probability of each hypothesis
Maximize ratio of the probabilities at each pixel location
Threshold to locate targets B A

15. Significance Verify algorithm for use in hyperspectral remote sensing
Prove algorithm’s potential application to areas other than remote sensing
Prove potential usage of hyperspectral imaging in document restoration
New tool for scholars
Locate spectral features useful in future research
Make data available to other researchers in the field of document restoration

16. Experimental Design and Methods
Verification of Algorithm
As described under specific aims - different images and targets
Concentrate on thresholds
Collection of ink spectra
Travel to Walters Art Gallery in Baltimore, MD
Tungsten-halogen illumination
Take direct reflectance measurements of Palimpsest with ASD
Image with CCD mounted with tunable filter ranging from nm

17. Experimental Design and Methods
Implementation of algorithm on palimpsest
Creation of invariant space
Method 1 - use MODTRAN to characterize illumination variation due to sensor/illuminant geometry, then find basis vectors
Method 2 - flat-field images to eliminate variations in illumination, use scene pixels to determine target basis vectors

18. Experimental Design and Methods
Implementation of algorithm on palimpsest
Search for underwriting ink
Method A - use underwriting ink as target
Method B - if under/overwriting inks are not spectrally distinct, use “ink” as target and subtract overwriting ink from final image
Method C - combine above with spatial methods of document restoration (approximate matched filters, etc.)

19. Resources and Environment
SUN workstations and existing images of Palimpsest
Analytical Spectral Device
Spectral range: nm, nm increments
Fiber optic cable used alone or with lenses to reduce field of view (from 5 degrees to 3 degrees or 1 degree)
X-Y translation table, CCD camera, liquid crystal tunable filter

21. This may change depending on arrangements with the museum and my availability.

22. For additional information: