SURF provides an efficient tool to perform SAR image geo

Report
An Efficient Automatic Geo-registration Technique
for High Resolution Spaceborne SAR Image Fusion
IGARSS 2011
28/July 2011
Woo-Kyung Lee and A.R.
Kim
Korea Aerospace University
[email protected]
Motivation
Simple approach to the SAR image registration and fusion
- As the resolution level improves,
* the unique feature of the radar imaging becomes prominent and the task of
image fusion with optical image becomes complicated,
* the number of pixels increases and the amount of resources for calculation
such as memory and time consumption escalates exponentially.
To relieve the burden of the work and make it done in real time.
Efficient image matching in both rural and urban regions
One click
Let the machine do the rest of the job
in almost real time
SAR Sensor and Geometric Characteristic
Side-looking Observation
-Image processing depends on the surface characteristics and structures
-Radar images suffer from unrealistic distortions
-Non-linear distortions along range, Shortening from shadow region
System error
- Inaccurate Doppler parameter estimation leads to geocoding errors
- Unstability in internal system clock and orbit parameters
SAR vs. optics images
Image acquisition
SAR Sensor and Geometric Characteristic
Source of SAR geocoding errors
Error Source
Effect of Error
SAR sensor
Effect of Error
- Electronic Time Delay
- Slant Range Error
- Range Location
- Incidence Angle Estimation
- Range Scale
- PRF Fluctuation
- Azimuth Scale
Earth
- Earth Rotation
- Side-looking
- Target Height
Platform
- Inclination Angle
- Yaw Angel Error
- Pitch Angle Error
Earth
- Azimuth Skew
- Range Non-Linearity
- Foreshortening, Layover,
Shadowing
Platform
- Image Orientation Error
- Squint Angle
- Doppler Centroid
Error correction method
- Geometric Calibration
- Deskew
- Ground Projection
- Image Rotation
- Terrain Correction
SAR Sensor and Geometric Characteristic
Geometrical distortions in SAR images
- Mismatch between SAR and Optical images
Optics
SAR
(a) Azimuth Distortion
(b) Non linear Range Error
(c) Deskew
SAR Geo-correction with satellite internal data
Ground projection example
Azimuth
-Slant range function has non linear scale variations
Slant range image
Range
Sland based
Ground range image
Reference image
(EO image)
Ground Based
-Discrepancy exist compared with the reference image
- Distortion between SAR and EO are case-sensitive
GCP based geometry registration
Basic Principle
- A reference image is chosen to be used for geometric correction or
fusion
- Multiple GCP(Ground Center Point)s are selected and directly applied
to individual position error calculation and correction i
- Based on the selected GCPs, image transforml function is
characterized that best describes the discrepancy between the images
- Original image is re-sampled and re-arranged by the generated
transform function
Choice of GCP
- To perform geometrical calibration and restore distortion, the GCPs
in the SAR images would be re-arranged into the true ground positions
- It becomes most essential to pick up the best candidates of GCPs
- In general, distinctive features such as road, building, bridges,
reflectors are chosen that are easily discriminated for convenience
- Manually? Or Automatically?? Who will chose what points??
Selection of GCPs within SAR images
Difficulty of GCP choice
- Speckle noise inherence in SAR image makes it difficult to guarantee
to pinpoint precise positions that correspond to the reference points.
- A human work of manual GCP selection is never reliable
- The number of available points are case-sensitive and still limited
by the existence of the distinctive features
- The precision of the GCP location is not fully guaranteed and the
error variance may increase in coarse resolution images.
Optical image GCP
SAR image GCP
Methodology
SURF algorithm
-Speeded-Up Robust Features (SURF) is known to have highly robust
performance
-Scale, rotation and illumination-invariant feature descriptor.
- Adaptive for noisy environment and mutll-scale images
- Only summing operation is involved in producing integral image to
match the scale and calculation is speeded up
Selection of GCP and matching
- Selection of matching points(GCP) are performed using feature
vectors described by Hessian Matrix
-The size of the constructed Hessian matix can be varied and can be
increased to multiple dimensions as desired
-The number of dimensions is limited by the complexity, time
consumption and precision of the image matching.- case sensitive
- Parameters required for the decision algorithms is set intuitively
- This work is motivated to find the decision parameters automatically
compromising the performance and the complexity
Block diagram for GCP pair selection
Integral image generation
-For the given image, an integral set of points are summed together
-The size is variable depending on the scale and complexity of the image
- Simple summations of intensity levels are performed over two
dimensional domain
: A +B + C + D
GCP candidate generation
- The Hessian matrix , corresponding to each pixel, is simplified by
summation with adjacent cells
- The image scale is varied and the simplified Hessian matrix is obtained
for each scale space
- Harr-wavelet responses are calculated and the feature descriptor is
generated
- The polarity of the image intensity variation is investigated and stored
Harr wavelet
X direction
Y direction
X, Y direction
GCP selection
Principle
- Find a pair of feature descriptors that best fit to each other
- One-by-one comparison is straightforward but time-consuming and
does not guarantee successful matching due to increased ambiguity
- Construct a look-up table for the feature descriptor
- Each feature descriptor is indexed depending on their size, variation rate,
orientation
- For a given GCP , a “search process” is performed within other look-up
table generated from reference image and the best matching pair is
selected
- Nearest neighbor search is adopted to find the correct matching pair
GCP pair selection
Define threshold level
- Among the selected GCPs, the Euclidian distance (x, y) – (x’, y’) are estimated to
find the nearest points with similar feature descriptor
-The rates of intensity variations along orientations (denoted as A and B) are
considered as weighting factors
- Distance multiplication is performed
-The number of orientation can be increased in order to reduce ambiguity and
avoid wrong decision.
- Appropriate threshold level is required to compare with the distance
multiplication and make a decision
-The GCP match is confirmed when the distance multiplication is less than the
threshold level
- Image Projective Transform function is deduced from the matching GCPs
Overall procedure diagram for image matching
GCP generation
GCP selection
Transform function
GCP extraction demonstration
GCP extraction from SAR images
-ScanSAR image is exposed to higher noise level and composed of
extended resolution cells.
-GCP candidates are extracted from both images using the same
Hessian matrix structure
- The number of GCP points appear to be close to each other despite
the gap in the image quality
GCP number
SURF
(a) Stripmap image
(b) Scan image
a
1870
b
1667
Geo-registration demonstration
-Original SAR image is corrected using GCP matching and transformation
-Strip mode SAR image over Vancouver, Canada is geo-registered using the
reference image in Radarsat-1 SSG format
- The threshold level is set to be zero for convenience
881
557
GCP # vs. Threshold level Time consumption vs Th. Level
Raw
GCP selection for raw image
Reference
912
544
GCP # vs. Threshold level Time consumption vs Th. Level
GCP selection
GCP selection for reference image
Image Matching Demonstration
-Original image is geo-corrected
Fusion image
Measure of registration errors
-RMSE(Root Mean Square Error) is calculated for the selected GCPs
Displacement Error in pixel
RMSE (x, y)
RMSE
(average)
Corrected
x
y
0.63
0.8
1.02
Reference
The average deviation is about one point pixel size
Application to higher resolution images
-Reference image of TerraSAR in EEC format, Toronto,
- The number of GCP increases consistently when the level of
correlation between the two images are high
1.72
1595
0.81
252
Original
As the similarity
of the images are
high, the GCP
increases
consistently as
the “Threshold
Level” decreases
Reference
3.21
2680
GCP variation rate
1.47
404
GCP selection
Image Matching Demonstration
After fusion
Mismatch Error Estimate
Displacement Error in pixel
RMSE (x, y)
RMSE
(average)
Corrected
x
y
0.7
0.32
0.77
Reference
The average position error is less than one pixel
The performance of the matched GCP selection is affected by the
image resolution
Mismatch error is reduced as the image resolution improves
Image fusion of SAR over EO
- This case is where SAR image constitute a small portion of the EO image
-GCPs from the two images are distinguished
- The matching GCPs are easily identified by the nearest search algorithm
Number of GCPs
(a) JERS SAR image
(b) LANDSAT EO image
a
375
b
1326
Automated threshold level selection
There exist a turning point, where
further reduction of the threshold level stops
affecting the number of GCP matching points
Computer traces the variation of
the available GCP matches and
find the turning point
- Threshold variance
Threshold 500, Matching image (14 points)
Threshold 350, Matching image (15 points)
Threshold 200, Matching image (15 points)
GCP matching and Image transformation
-Matching GCP selection process stops automatically
and image transform function is obtained from the
selected points
 15 points extraction
Feature points extraction
Find equation
Transform equation
Source image
Reference image
x
y
x
y
139
88
74
247
119
143
359
37
459
175
94
203
254
232
137
234
350
268
365
366
509
294
514
425
192
402
231
209
503
160
137
234
236
221
94
203
362
272
390
386
266
224
173
218
280
283
177
372
349
250
365
329
440
296
87
391
383
237
467
270
Result of the geo-registration
Overlaid image
-RMSE is affected by the resolution discrepancy and inherent image property.
- Here it is given as 1.26 pixel
Automated Geo-registration software
-Usually, the threshold level is manually set-up by user looking into
the complexity of the images and resultant fusion performance
- This procedure is replaced by compute search algorithm, where the
threshold level is traced to find the turning point
- Total elapsed time is within several minutes and will be further
reduced by adaptive search algorithm
Original
Reference
GCP selection and
matching
Corrected
Fusion
Performance analysis vs. Resolution
- The number of total GCP is not affected by modes(Stripmap and Scan)
GCP number
SURF
(a) Stripmap image
a
1870
b
1667
(b) Scan image
- However, the RMSE is measured as 0.94 for
ScanSAR mode while it is 0.34 for stripmap
mode
- It appears that the performance improves as
the resolution improves
Insufficient information for SAR geometry
Limited information
-Internal data within SAR instrument fails to retrieve shadow region
-There is non-linear discrepancy between slant and ground ranges
-Generate Errors in geometrical coordinate
-Need external references to retrieve broken information and correct
errors in ground range allocations
- foreshortening, layover, shadowing
Foreshortening
Layover
Shadowing
Impact of the ground characteristics
- Diverse ground geometry becomes a source of matching errors
- Mountain areas are severely distorted from the EO case
- Need to adopt separate transform functions within the image
Coast line area
Coast line fusion
Mountain area
After correction
After correction
Mountain area fusion
Matching Performance Comparison
- Image distortion is not compensated for by the simple GCP transformation
- Need to divide blocks and adopt modified transform functions separately
Error
RMSE (x,
y)
X
Y
0.33
0.26
RMSE
0.42
Coast line
Error
RMSE (x,
y)
RMSE
Mountain
X
Y
1.46
1.73
2.27
Modified Transform functions
- Image is divided into blocks according to the geographical properties
Errors
RMSE (x,
y)
x
y
1.35
1.2
RMSE
1.8
Average 1.8 RMSE error
Mountain Area
Errors
RMSE (x,
y)
RMSE
x
y
0.6
0.3
0.67
Average 0.67 RMSE Error
Urban Area
Application of separate transform function leads to the reduction of RMSE
Summary
Conclusion
- SURF provides an efficient tool to perform SAR image geo-registration
- A choice of threshold level is required to perform efficient of GCP matching and
it can be automated by tracing its variation curve
- The image matching algorithm works with various SAR and EO images and the
average RMSE is measured to be around 1 pixel.
- Image blocks containing mountain areas need separate GCP matching and
transform function to compensate for image distortion
Further work
- Need to develop optimized transfer function for different type of ground
characteristics
- Indexing of GCP is performed based on their intensity and variation vector
-With the introduction of adaptive indexing table for selected GCP, the automated
image matching is expected to be carried out in real time
-

similar documents