Graphics Reference
In-Depth Information
by utilizing robust image registration and motion compensation to align successive
image frames that may have random inter-frame motion due to camera movement
and/or object motion. In our CASSI system if the captured images are well aligned,
then averaging N images will not only reduce blur, but will also increase the SNR
by a factor of N as compared to a single capture with limited exposure time.
Image registration (e.g., motion estimation) methods can be broadly grouped into:
feature based motion estimation and global motion estimation. Feature based meth-
ods [
29
,
30
] use features extracted by image processing (corners, oriented edges, and
so on) to find correspondences in successive images and hence solve for a motion
model between these images. Global motion estimation [
31
] uses the intensity infor-
mation from all the pixels in the image to directly recover the motion between suc-
cessive image frames.
In our CASSI system the coded aperture (CA) introduces a fixed pattern in the
image (before datacube reconstruction) which masks any structure of objects in the
scene. Hence it is impossible to extract features from the scene that are suitable for
motion estimation. The global methods, on the other hand, use all the pixels in the
image. If the effect of the CA pattern can be averaged out and only the contributions
from the signals in the scene remain, then this method would be a solution to the
problem.We choose this approach sincewe can compensate for the knownCApattern
in every captured frame while maintaining a reasonable SNR by normalization.
Our choice for global motion estimation method is the hierarchical model-based
motion estimation [
31
] for which SRI has a real-time implementation. The estimation
process involves Sum-of-Squared-Difference (SSD) minimization using a motion
model between image frames, with Gauss-Newton minimization employed in a final
refinement process.
13.2.3 CASSI Based Spectral Classification
An experiment was designed to demonstrate the CASSI-based spectral classification.
A CASSI system (Fig.
13.4
a) was used to measure the fluorescent visible colors
illuminated by an UV light source. Four highlighter colors were used to generate
green, pink, orange, and yellow alphabet letters in Fig.
13.4
b. A single frame of
CASSI measurement was shown in Fig.
13.4
c for the datacube reconstruction. In
the measurement, the “G” letter at the second row did not interact with the UV
illumination because the letter was marked by a non-fluorescent green Sharpie pen.
Two spectral channels of the CASSI reconstruction are compared at Channel 12 and
18 (Fig.
13.4
d, e) from 20 channels. At Channel 12 the “G” letter is dominant and at
Channel 18 the “O” and “?” letters are dominant. The question mark “?” was added
in the object scene to demonstrate a multi-target classification. The question mark
has the same spectrum with the “O” letter because “?” was marked by the orange
highlighter color. In the datacube, the “G”, “P”, “O”, and “Y” letters have different
intensity distributions along the spectral channels.
