Designing Vision Systems that See Better - Advances in Embedded Computer Vision - page 260

task-specific information [ 22 ] to measure projections for either reconstruction
[ 17 , 19 , 20 , 23 ] or classification [ 15 , 16 , 18 ]. Since the aim is to perform high-
level computer vision tasks directly on compressed measurements, such systems can
dramatically reduce the amount of data for a given SNR compared to traditional
imaging architectures.
One major limitation of feature-specific imaging, however, is that the measured
data is only useful for a specific task at any givenmoment. Thus it is nearly impossible
to reuse the data for any other purposes such as updating statistical models of objects.
This limitation is especially severe in spatial dimensions since many scene structures
will be lost. What is needed is a unified, extensible framework for task-specific imag-
ing systems that integrally performs detection, classification, tracking, and learning
directly on compressive measurements . This approach is favorable because we are
no longer dependent on the quality of the decoded imagery.
In the rest of this chapter, wewill describe two examples of computational imaging
solutions to motivate the application areas and to reinforce the advantages of this
research area. Then we present the architectural framework of an adaptive task-
specific imaging system, optimized for the embedded vision task at hand.
13.2 Multi Spectral Coded Aperture Camera
Multispectral imaging (MSI) enables recovery of spectral properties of material in
a scene, based on the material's spectral responses to current illumination (e.g.,
reflection, absorption, fluorescence). In this section, we describe a key development
towards a small form-factor imaging spectrometer that can enable instantaneous cap-
ture and analysis of the spectral signatures of all objects in the scene. This spectral
classification system is achieved by combining a coded aperture snapshot spectral
imager (CASSI) with a multispectral detection algorithm. We further improve the
signal to noise ratio with temporal analysis enabled with image registration of cap-
tured images. The spectral datacube is captured and encoded simultaneously into a
2D image using a code aperture, and later decoded by sparsity-based computational
framework. An adaptive-cosine estimator (ACE) is used to quantitatively detect and
classify the target objects from the decoded spectral cube [ 24 , 25 ]. We present details
about our optical design, algorithm, and selected results of our camera system.
13.2.1 Computational and Optical Co-design
CASSI was first presented in [ 26 ] for a snapshot MSI. The optical encoding with
a static coded aperture and a dispersive element enables the system to compres-
sively record the spectral datacube and the computational decoding reconstructs the
datacube from the compressed measurement. With respect to computational imag-
ing, the optical measurement is formulated by a linear model and the input data is