The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part BI. Beijing 2008
1359
proposed algorithm, every object in an image is encapsulated by
a tree structure, where root of the object tree is the brightest
pixel in the object and leaves are the pixels at
object/background boundaries. The pixels belonging to the
object but lying between the root and the leaves are non
terminal nodes of the object tree. Many object trees are thus
created all over the image corresponding to various objects. All
object trees whose root gray value is greater than or equal to
1000 are identified as candidates for smoothing. A mask
identifying pixels corresponding to such object trees
(comprising of root, nodes and leaves) is generated. Smoothing
(3x3 averaging) is done only for the masked pixels in the image.
Such a selectively smoothed image forms the input for MTF
correction. Due to prior smoothing of bright objects, ringing
artifacts do not arise in the restored image.
4.1.3 Artifact Removal Post-processing:
Post-processing involves spatial neighborhood operation on the
restored imagery to remove false shadows/dark spots. Original
un-restored image and restored image form input to the
proposed algorithm. All pixels in the restored image, where
ratio of restored and unrestored values exceeds a threshold (say
2.0) are firstly identified. The intensity of such pixels is
recomputed using corresponding un-restored gray value from
original image and restored gray value from the restored image.
4.1.4 Wavelet based denoising:
The imaging mechanism (in this case, pushbroom scanning by
an imaging sensor mounted on a satellite) acts as a low-pass
filter and introduces degradation and noise in the observed
image. The degradation is less for the lower spatial frequencies
as compared to the mid and high spatial frequencies. As the
high spatial frequencies (eg. features like edges) suffer more
degradation (appear blurred), the signal power at high
frequencies is low. However, the noise power is uniform at all
frequencies (such type of noise is termed as ‘white’ noise).
Therefore, at high frequencies, noise dominates over the signal.
Restoration filter achieves deblurring by amplifying the high
frequencies. The gain of the restoration filter is adjusted such
that it is high at mid and high frequencies and unity at low
frequencies. As a result, after restoration, the image appears to
be deblurred (sharp), but noisy. The nature of noise that was
‘white’ prior to restoration is also changed to ‘colored’ during
the process due to unequal gains of restoration filter. The
amplified noise is particularly noticed in the uniform areas of
the image.
The stagger removed, MTF corrected (restored) and artifact
removed image is decomposed into complex sub-bands using
the dual tree complex wavelet transform. The signal is assumed
to be concentrated on the larger scale coefficients, while noise is
distributed with same variance over all the coefficients. As the
signal and noise are located in nearly separate sub bands, the
next step is to efficiently estimate the threshold that can
eliminate the noisy coefficients. The thresholds are calculated
using the adaptive ‘shrink’ method. The new wavelet
coefficients are calculated by applying soft thresholding on the
sub bands. Inverse transform of the thresholded bands gives the
denoised image. The complex dual tree discrete wavelet
transform is very efficient in higher noise conditions.
4.2 Restoration of Cartosat-1 Imagery:
The laboratory measured point spread functions for the Fore and
Aft sensors of Cartosat-1 were taken as the degradation function
for restoration of respective imagery. Wiener filter, which
incorporates the degradation function as well as the model of
noise, was designed in frequency domain for each sensor. After
several experiments and independent qualitative and
quantitative evaluations, the noise to signal ratio (nsr) was
modeled to vary exponentially between nsrmin and nsrmax for
low to high frequencies. The values of nsrmin and nsrmax were
tuned suitably to control sharpness and noise.
Restoration of the fore and aft images was performed block-
wise. Every block of the input image was filtered with the
Wiener filter in the frequency domain. Overlap of few pixels
was maintained between successive blocks to avoid artifacts at
block boundaries. Fast Fourier Transform (FFT) techniques
were used to achieve high speed. The results of restoration are
shown in Figures 3-6. Subsequent to restoration, Wavelet based
denoising was performed.
4.3 Evaluation:
Apart from visual improvements (see Figures 3-6), the restored
and denoised Cartosat-1 stereo images were independently
evaluated quantitatively for their performance improvement in
DEM generation. DEMs were generated from original stereo
orthokit products as well as MTF enhanced products and the
resulting DEMs were compared with high accuracy reference
DEMs provided by respective Principal Investigators for two
test sites, namely Hobart in Australia and Castel Gandolfo near
Rome in Italy. The results are shown in Tables 1 and 2 and
Figures 1 and 2. It could be observed in Tables 1 and 2 the third
column values are consistently found larger than the
corresponding values in column 2 indicating the enhanced
performance in DEM generation due to the MTF improvements
carried out on the stereo images. DEM improvements have
expectedly been resulted due to more accurate stereo image
matching after MTF enhancements. To observe the visual
quality improvements in Figures 3 to 6, readers are
recommended to observe them in softcopy version of this Paper
accessible from ISPRS web site.
Height
Difference
Comparison of DEM
generated from
Original Orthokit
with Reference DEM
Cumulative
Percentage
Comparison of DEM
generated from MTF
Enhanced Orthokit
with Reference DEM
Cumulative
Percentage
Up to 1 m
37.5
43.25
Up to 3 m
55.16
59.98
Up to 5 m
70.03
70.74
Up to 10 m
89.76
91.65
Table-1 : DEM Results for Hobart Test Site
Height
Difference
Comparison of DEM
generated from
Original Orthokit
with Reference DEM
Cumulative
Percentage
Comparison of DEM
generated from MTF
Enhanced Orthokit
with Reference DEM
Cumulative
Percentage
Up to 1 m
18.39
18.58
Up to 3 m
58.67
62.15
Up to 5 m
75.81
78.88
Up to 10 m
84.97
93.12
Table-2: DEM Test Results for Castel Gandolfo Test Site