My
current research interests occupy
the intersection of physical optics and signal
processing,
with an
emphasis on the
use of structured illumination. But,
in previous years I have worked on a broad array of topics. A
brief selection of past and present efforts are
summarized in this page. I encourage you to
navigate the tabs on this page for more
information.
-
Optical Imaging
-
Computer Vision
-
Image Processing
-
Awards
The
acquisition and sensing mechanism in existing electro-optic imagers is
optimized
for line of sight operation. As a result, imagers cannot discern objects that are beyond
the line of sight and potentially hidden from
view. The OMNISCIENT program seeks to transform image sensing
by
exploiting the wave nature of light and the availability of scattering
surfaces, in identifying a
latent image of the unobservable objects.
This research effort aims to build imagers whose
functionality may be rapidly adapted to suit
evolving tactical needs. Among other things, these imagers are expected
to support target/threat detection and identification at increased
standoff, in a wide field of regard, with the highest clarity, through
the cover of darkness and obscurants. The development of such imagers
is facilitated by two recent innovations
- Emergence of computational imagers that afford novel imaging capabilities by manipulating the light distribution in the object and/or image volumes. Examples of novel capabilities includes simultaneous super-resolved imaging and ranging.
- Development of engineered optical beams that can squeeze light into regions smaller than the diffraction limit
With
the aid of DoD funding, my group at SMU has
developed computational imagers whose resolving power is fully
decoupled from the constraints
imposed by the collection optics (such as diffraction and aberrations)
and the detector (varying degrees of aliasing ranging from single
photodiode to focal plane array). Additionally,
these imagers feature support for point spread function engineering,
foveated imaging and ranging. Such
disparate capabilities are realized by processing images acquired under
spatially patterned illumination. The
Moiré fringes arising from the heterodyning of object detail and the
illumination pattern, encapsulate spatial frequencies lost to
diffraction. The deformations in the phase of the detected illumination
pattern, encode range information.
This
work has been recognized with multiple awards. More
information on the research effort is available here.
The Spatial Frequency Response (SFR) of a digital
image acquisition system neatly encapsulates the influence of optical
elements, pixel MTF and camera electronics, on image
quality. The slanted-edge algorithm outlined in the ISO12233 standard
is the gold standard for identifying the SFR. Critical examination
of the slanted-edge method for multispectral cameras revealed
inaccuracies in the estimated SFR, on account of demosaicing. The
objective of this research effort was to resolve inaccuracies in the
estimated SFR, by eliminating the need for demosaicing. The task is
accomplished by accommodating the sparse sampling structure of the
Bayer color filter array (CFA) within the slanted edge method. The
method facilitates the comparison of the image quality of cameras with
vastly differing CFA architectures, and help characterize the impact of
a specific demosaicing algorithm on image quality.
Additional
information
is available here.
The
wealth of literature on the topic of Digital Super Resolution fail to
provide satisfactory answers to simple questions such as
Additional
information
is available here.
- Is perfect recovery of an optically band-limited image possible under precise knowledge of translational motion ?
- What are preferred optical PSF's (and focal plane masks) for digital super-resolution ?
- Under what conditions can we disregard the use of regularization ?
Accurate
measurement
of the physical distance to
an object (ranging) has numerous applications in areas ranging from
metrology to logistics. The
commercial availability of photonic mixer devices has made it
increasingly simple to acquire
range information using a temporally modulated light source. The time
difference between the incident and reflected light paths determines
the distance to each scene point. The mechanics of ranging using the
above notion is predicated on the
availability of a single-bounce light path from the illumination source
to individual sensor pixels. However, the view is inconsistent with
the infinitely many ways in which light is transported
from the source to the detector. The result is measurement errors in
the estimated scene geometry.
The problem has gained tremendous attention in the vision community and a variety of inreasingly sophisticated solutions have emerged. Our exceedingly simple solution to the problem is inspired by point-scanning LiDAR. We seek to replace the flood-illumination module in commodity ToF sensors with a fan-out of spatially confined beams. A range map of the scene is assembled by consolidating range measurements that are acquired as the illumination pattern scans the scene.
The problem has gained tremendous attention in the vision community and a variety of inreasingly sophisticated solutions have emerged. Our exceedingly simple solution to the problem is inspired by point-scanning LiDAR. We seek to replace the flood-illumination module in commodity ToF sensors with a fan-out of spatially confined beams. A range map of the scene is assembled by consolidating range measurements that are acquired as the illumination pattern scans the scene.
This research effort dispels popular notions
concerning the use of Least Squares estimators (LSE) in computer vision
(CV). It is widely believed that coordinate normalization is mandatory
in LS parameter estimation, and that the heteroscedastic nature of LSE
in CV induces a larger bias in LSE compared to maximum likelihood
estimation
(MLE). The first notion is dispelled by extending the notion of
coordinate system invariance in LS curve fitting, to parameter
estimation in computer vision. This reveals the existence of a family
of LS estimators that are invariant to coordinate normalization and
consequently do not need coordinate normalization. The second notion is
dispelled by using higher-order perturbation analysis of linear
systems, to demonstrate the existence of LS methods with a smaller bias
compared to the ML estimator.
Additional
information
is available in these publications: Paper1,
Paper2,
Paper3,
Paper4,
Paper5
This work examines the problem of
removing occluders in an image using inpainting. We develop a geometric
method that utilizes a second image of the scene from a different
viewpoint, to inpaint the occluded objects. We recover the missing
intensities by using the geometric relationship between corresponding
points in the two images. The relationship is generally specified by
the "epipolar line constraint", and degenerates to a projective
homography under special circumstances. We fill-in missing pixels by
copying information from the respective epipolar lines in the second
image. The success of the approach hinges on the ability to estimate
the "epipolar line constraint" from noisy correspondences. To this end,
we analyze the uncertainty in estimating a homography from noisy
correspondences. We rely on the knowledge of this uncertainty to
identify parallax vectors best suited for estimating the epipolar
geometry. We do not make any explicit assumptions about the nature or
the extent of camera motion, only requiring that the occluded objects
are static and undergo limited perspective change.
Additional
information
is available here.
More information on this inter-disciplinary
research effort will be made available soon.
Traditional imaging sensors struggle to capture
the wide dynamic range of natural scenes with exceedingly high
contrast. The result is saturation and underexposure of the captured
images. The aim of this research effort was to build a computational
model for contrast perception that reduces the wide dynamic range in a
scene, while still preserving the details. The idea is to preserve
local contrast variations (reflectance) and attenuate the intensity
(illumination) in areas of high-contrast. The success of the technique
hinges on the ability to factorize the captured image into the
illumination and reflectance components. An estimate of the
illumination component is obtained by solving a modified heat equation
that is initialized with the captured image field.
Additional
information
is available here: Paper,
Slide deck
This work explores the use of standard adaptive
filtering techniques in suppressing broadcast logo's embedded in
television material. The real time processing requirements of the
problem necessitate the use of adaptive filters with low computational
burden. The proposed approach uses a simplified version of the standard
adaptive sinusoidal interference canceller, in an attempt to suppress
the logo at the pixel level. An array of these pixel level logo
suppressors is used to suppress the complete logo.
Additional
information
is available here.
One of 54 early career researchers recognized as up-and-coming standouts in their fields, capable of discovering and leveraging innovative opportunities for technological surprise. The research effort recognized in this award is described here.
AcknowledgmentI would like to acknowledge the help of the following individuals in shaping my research efforts: Prof. Marc Christensen, Prof. Panos Papamichalis, Dr. Predrag Milojkovic, Prof.Kenichi Kanatani, Indranil Sinharoy, Dr.Vikrant Bhakta and Prof. Dinesh Rajan. |