Graphics Reference
In-Depth Information
13.1.3 Processing
With new mapping of how light is captured and coded through the optics and sensor,
and also through various illumination techniques, a decoding step is required to
“re-create” the image from raw sensor data. Very often, compressive sensing (CS for
short) decoding algorithms provide a formalism in which a physical sampled data set
is analyzed to generate data in a different domain. For example, 3D spectral datacube
is physically mapped to a 2D focal plane sensor, in which the decoding process will
recreate the 3D datacube with certain fidelity [
6
].
Efficient CS decoding algorithms will take advantage of the sparse nature of the
mapping matrix that is used to translate raw sensed data to the final domain. CS
decoding algorithms fall into two main categories:
convex optimization
and
greedy
algorithms. The former cast the problem into a linear program and apply efficient
algorithms to its solution; the latter refine an initial estimate one element at a time.
Compared to greedy algorithms, convex optimization algorithms tend to produce
more accurate results but require higher computational complexity. Examples of
two convex optimization algorithms that are reasonably efficient and produce good
results for datacube reconstruction are TVAL3 [
12
] and TwIST [
13
]. TVAL3 is a
decoding algorithm that can take advantage of a sparse sensing matrix and a signal
that is dense in its native sampling domain. It is also optimized to work with
images
,
since it implicitly performs Total Variation (TV) minimization—it searches for a
solution with a sparse 2D
gradient
. TwIST is another decoding algorithm that has
been optimized for image reconstruction. It is somewhat more flexible than TVAL3
but can take slightly more time to produce a result. Like TVAL3, it is a convex
optimization algorithm.
13.1.4 Putting Everything Together
There is not a common camera architecture for computational imaging as there is
much research left to explore before arriving at any optimal configuration of optics,
sensor and processor. While there are camera platforms such as the Frankencamera
[
14
] that exposes internal processing chain to allow for developers to create appli-
cations in this area, the field is a wide research territory to arrive at different con-
figurations suitable for many embedded vision application domains, including face
recognition [
15
], multivariate optical computing in spectroscopy [
16
], image sensing
[
17
], material classification [
18
], and compressive video reconstruction [
19
].
Specific to the discussion about optimization for computational imaging, we advo-
cate an alternative approach to directly capture only useful
features
(e.g., some key
spectral bands) that are critical to perform a specified task. Known as
feature-specific
imaging
[
20
,
21
], this technique employs novel optical modulators to measure
lin-
ear projections
of incoming radiance. Depending on the nature of the given tasks,
such computational imaging systems can be optimally designed by maximizing
