Sixth Annual Army Landmine Research Technical Review Meeting

1
Sixth Annual Army Landmine Research Technical
Review Meeting
IMAGE PHASE-ONLY INFORMATION FOR LANDMINE
CLASSIFICATION USING ANN and DT/WAVELET
SUPERRESOLUTION FROM IMAGE SEQUENCE
PI DR. N. K. BOSE
THE PENNSYLVANIA STATE UNIVERSITY UNIVERSITY
PARK, PA 16802
DATE THURSDAY JANUARY 23, 2003 PLACE
SPRINGFIELD, VA. SPONSOR DR. WILLIAM A. SANDER
U. S. ARMY RESEARCH OFFICE NORTH CAROLINA
277-09-2211
GRANT DAAD 19-00-1-0539
2
The Importance of Phase in the Presence of Noise
for Image Reconstruction
Objective
To evaluate the importance of phase information
in the presence
of noise in image reconstruction.
Result
Reconstructed image from phase only information
has many important
features in common with the original image. Also,
reconstruction from noisy phase (and original
magnitude) is more robust to error than
reconstruction from noisy magnitude (and original
phase). Therefore, phase only information can be
used for various applications like signal
reconstruction, landmine classifier, and so on.
Original image and reconstructed images
Original image
Reconstructed image from magnitude with zero
phase by IDFT
Reconstructed image from phase with ensemble
magnitude by IDFT
Reconstructed image from phase with unit
magnitude by IDFT
Reconstructed image from noisy phase and
noisy-free magnitude by IDFT
Reconstructed image from noisy magnitude and
noisy-free phase by IDFT
3
Neural Network for Object Classification from
Phase Only Information
Objective
To assess the importance of noise-corrupted,
degraded and partial
phase information in object classification by
artificial neural network (ANN) in applications
like landmine identification and classification.
Result
Phase information is more error-tolerant than
magnitude information
in terms of signal reconstruction. The number of
input neurons can be reduced to adequate size to
simplify ANN and decrease ANN training time.
Therefore, partial phase only information can be
used as input data in object classification by ANN
Overall neural network for object classification
4
Delaunay Triangulation based High Resolution
(DTHR) Algorithm
Surface Approximation
DT constructed from first 3 LR frames

• The smooth surface, used in the research at
  present, is defined by the bivariate polynomial
  (splines also usable)

• By substituting the value in each vertex i (i
  1,2,3) along with corresponding estimated
  gradients, one gets

i 1, 2, 3.
Updated HR image after adding frame 4,5, and 6
MSE 49.43
Initial HR image from first 3 frames MSE 85.27
Sample LR frame
Original HR image
vs
Reference S. Lertrattanapanich and N. K. Bose,
"High Resolution Image Formation from Low
Resolution Frames Using Delaunay Triangulation",
IEEE Transactions on Image Processing, vol. 11,
no. 12, December 2002, scheduled to appear.
5
Postprocessing Processes
Multiframe Noise Filtering
Richardson-Lucy (R-L) Algorithm

• The R-L algorithm is a maximum-likelihood (ML)
  method for image restoration when the complete
  knowledge of PSF is known.
• The R-L iteration is

• Shepp and Vardi (1982) derived the R-L algorithm
  as a special case of the expectation maximization
  (EM) algorithm of Dempster et al (1977) which has
  been proved to converge to the ML solution.
• In addition, it was also proved to converge at
  least in the weak sense that is, iteration never
  cause the likelihood to decrease Hanisch et al
  1997.

• The PSNR is increased from 28.71 to 31.03 dB.

Averaged image
Multiframe Iterative Blur Identification
g1
g2
g3
g4

• The iterative blind deconvolution (IBD)
  algorithm Biggs and Andrews 1998 involve the
  alternate application of R-L algorithm to image
  and PSF estimates until both converge.
• The single frame IBD algorithm can be directly
  generalized to multiframe case. Wraparound
  artifacts and noise amplification controllable.

PSF estimate
6
High Resolution (HR) Regions of Interest (ROIs)
from Real Video
Panoramic image (video mosaic) constructed from a
real video sequence (eo_bldg2.AVI) supplied by
the Air Force Research Laboratory (AFRL)

• Click the image below to show a real video
  sequence.

Reference S. Lertrattanapanich and N. K. Bose,
"Latest Results on High-Resolution Reconstruction
from Video Sequences", Technical Report of IEICE,
DSP-140, The Institution of Electronic,
Information and Communication Engineers, pp.
59-65, December 1999, Japan.
Shadows are clearly resolved here.
Helicopter landing pad
There are 4 windows here.
Electric pole