Specifications

Submitted to Studies in Conservation, March 2006
5
CCD sensor, filters had to be defined and fabricated. An optimization was performed to select
filters from among the Schott filter glass catalog. The optimization considered spectral and
colorimetric accuracy, image noise, capture time, and fabrication simplicity and cost [9 – 11].
There were a number of different filter pairs with similar performance. We selected a pair in
which one of the filters resulted in spectral sensitivities similar to the sensor with its original
blue-green cover glass. Thus, the camera could also be used in the usual fashion as a color-
managed RGB digital camera. The final filters are listed in Table 1 and plotted in Figure 2. Each
was a “sandwich” of an absorption filter and a visible bandpass filter (i.e., a UV and NIR
blocking filter). The bandpass filter transmitted radiation between 380 and 750 nm. The surface
of each filter facing the sensor was anti-reflection coated to minimize inter-reflections between
the lens filter and CCD cover glass. The spectral sensitivities of the CFA sensor are plotted in
Figure 3. In the visible region, there are the expected three peaks at red, green, and blue
wavelengths. The red channel has sensitivity in the red and near infrared regions while the blue
and green channels have sensitivity in the near infrared. The effect of placing each optimized
filter in the optical path is plotted in Figure 4. The blue-green filter predominantly affects the red
channel sensitivity by defining the wavelength of peak sensitivity and bandwidth. The yellow
filter predominantly affects the blue channel sensitivity by narrowing bandwidth. The visible
bandpass filter, common to both filter sandwiches, limits spectral sensitivity to the visible region,
critical when correlating with the human visual system. Based on Figure 4, it appears that the
two filters have a minimal affect on increasing spectral information beyond that normally
captured with a CFA sensor. As described above, spectral reflectance factor is estimated by a
linear calibration transformation. This is equivalent to creating new spectral sensitivities by
weighted addition or subtraction of the spectral sensitivities plotted in Figure 4. As an example,
in Figure 5, the yellow-filtered blue channel was subtracted from the blue-green-filtered blue
channel and blue-green filtered red channel was subtracted from the yellow-filtered red channel,
plotted as dashed lines, along with the blue-green filtered spectral sensitivities, plotted as solid
lines. It is observed that the visible spectrum is sampled in five discrete locations. The two-filter
approach along with the calibration transformation has enabled the increase in sampling number.
Typically, the matrix coefficients are optimized to minimize either spectral or
colorimetric error. We have developed a technique where both errors were minimized [12],