Automatic Registration¶
Introduction¶
Automatic registration is the attempt to match a pattern in an image cube. Pattern matching has many different purposes.
Examples:
-
Register an entire image to a second image. The registration of an image is performed by moving (rubber-sheet) the pixels to output pixel locations that result in pattern matching directly to a second 'fixed' or 'truth' image.
- Use: coreg
-
Relative cartographic registration between a number of overlapping images. Pattern matching is used to build a control point network across a number of overlapping images for solving and adjusting the camera pointing that are then applied to map project the images.
-
Find and record the pixel location(s) of a camera reference mark (e.g., Vidicon reseau mark) across an image
- Use: findrx
Match Algorithm Definition Files (PVL)¶
Registration of images can be an advanced and complicated process. ISIS3 is designed to supply a choice of match/registration algorithms to handle a variety of image data conditions. An ISIS3 "plugin" algorithm expects all match algorithms to supply parameter information through a Parameter Value Language (PVL) definition file. Examples of match algorithm PVL's can be found in
$ISIS3DATA/base/templates/autoreg/coreg*.def
Registration applications (such as coreg) will allow the user to choose from this collection of PVL files in the user interface gui.
A PVL file can be generated for a selected match algorithm with autoregtemplate.
Match Algorithms¶
Maximum Correlation¶
The Maximum Correlation match algorithm is most commonly used. The
algorithm computes a
correlation coefficient that will range from no correlation (zero) to
perfect correlation (one).
Example templates for Maximum Correlation can be found in:
$ISIS3DATA/base/templates/autoreg/coreg.maxcor.p*.s*.def.
An example PVL for the Maximum Correlation follows:
Object = AutoRegistration
Group = Algorithm
Name = MaximumCorrelation
Tolerance = 0.7
End_Group
End_Object
Tip
A successful correlation is met if the Goodness of Fit is Greater Than the Tolerance parameter value for the Maximum Correlation Algorithm.
Minimum Difference¶
The Minimum Difference match algorithm performs a subtraction of the Pattern Chip and the sub-region of the Search Chip. The Goodness of Fit will range from zero for a perfect match, while larger values indicate a less likely match. The Goodness of Fit is measured as an absolute value and will never be negative.
Example templates for Minimum Difference can be found in:
$ISIS3DATA/base/templates/autoreg/coreg.mindif.p*.s*.def.
An example PVL for the Minimum Difference follows:
Object = AutoRegistration
Group = Algorithm
Name = MinimumDifference
Tolerance = 100
End_Group
End_Object
Tip
A successful correlation is met if the Goodness of Fit is Less Than the Tolerance parameter value for the Minimum Difference Algorithm.
Tolerance¶
The Tolerance parameter is used by both Maximum Correlation and Minimum Difference. Each correlation coefficient result (Goodness of Fit) is tested against the Tolerance. Tolerance is set within the PVL file along with the chosen registration algorithm.
Points that fail the Tolerance test are reported by the autoreg applications as:
-
FitChipToleranceNotMet
-
SurfaceModelToleranceNotMet
Chips¶
There are two chips used in automatic registration, the Pattern Chip and the Search Chip. The two chips are nothing more than sub-areas of the input image cubes, generally small in size. An NxM Chip is defined to be an N-Sample by M-Line region of an image. Basic elements of a chip are:
- N and M are natural numbers (1, 2, 3...,50)
- Like image cubes, chip coordinates are sample/line and 1-band based
- The center of the chips is (N - 1)/2 + 1 and (M - 1)/2 + 1
Pattern Chip¶
A Pattern Chip will contain the data you would like to match. This Pattern Chip is basically extracted from the 'Reference' image (it is also referred to as the 'Truth' image). The Pattern Chip is then walked across the Search Chip that has been extracted from the other image that is to be modified.
The PVL for a Pattern Chip defined as 25 Samples x 25 Lines (625 pixels) is:
Object = AutoRegistration
Group = PatternChip
Samples = 25
Lines = 25
End_Group
End_Object
Pattern Chip Size Requirement:
N + M >= 3. This ensures that the pattern is not a single pixel.
Tips
- The feature to match included in the Pattern Chip does not have to be in the center.
- The more "bland" the data is within the input images, the larger the Pattern Chip box should be. A larger box will allow for more pixel DN variance.
- In general, avoid very small box sizes less than 11 Samples X 11 Lines
- The smaller the size of the Pattern Chip, the greater the chance of matching too many areas in the search chip and returning a 'false positive' match.
Search Chip¶
A Search Chip is the sub-area of the image cube that the pattern might be found in. The Pattern Chip is walked through the Search Chip looking for the best match. The Search Chip must be larger than the Pattern Chip (at least one pixel bigger on line/sample size).
The PVL for a Search Chip defined as 45 Samples x 45 Lines (2025 pixels) is:
Object = AutoRegistration
Group = PatternChip
Samples = 25
Lines = 25
End_Group
Group = SearchChip
Samples = 45
Lines = 45
End_Group
End_Object
Search Chip Size Requirement:
N(search) >= N(pattern) + 2 and similarly for M. This ensures that the Pattern Chip spans at least a 3x3 window within the Search Chip. An important requirement for surface fitting in order to compute sub-pixel accuracy. (Refer to: Sub-Pixel Positioning).
Tips
- The Search Chip must be larger than the Pattern Chip (at least one pixel bigger on line/sample size).
- The Search Chip needs to be large enough to allow a 'best guess' or
predicted line and sample mis-registration offset between the images
that are to be correlated.
- SearchChipSize = PatternChipSize + (OffsetPixelEstimate * 2)
- BUT, the Search Chip shouldn't be too much larger than the Pattern
Chip.
- The difference in size could greatly affect how long the registration application would run.
- Refer to PVL example above:
- An estimated offset between two input images of 10 Lines by 10 Samples will need a Search Chip size (10 Line offset * 2) by (10 Sample offset * 2). The Search Chip size in this example needs to be at least 20*20 Lines/Samples bigger than the pattern chip.
- PatternChipSize = 25 * 25
- SearchLineDimention = 25 + (10 * 2) = 45 * SearchSampleDimention = 25 + (10 * 2) = 45
- SearchChipSize= 45 * 45
Restricting Input Pixel Ranges¶
The validity of the input pixels of the Pattern Chip is the very first test performed. Prior to the match algorithm being invoked during the walk process, a simple test is performed to ensure there are enough pixels to work with. Pixels are deemed valid if they are in the minimum/maximum range and/or they are not special pixel values.
The Pattern Chip is only checked once. If it does not contain enough valid pixels, the match is deemed to fail. Otherwise, the walk through occurs and the sub-area is extracted from the Search Chip. If this sub-area does not have enough valid pixels a match will be deemed to fail at that search location.
Points that fail the default or specified Valid Minimum/Maximum and/or ValidPercent tests are reported by the autoreg applications as:
-
PatternNotEnoughValidData
-
FitChipNoData
-
SurfaceModelToleranceNotMet
Valid Minimum/Maximum¶
Input image pixels (DN) values may be excluded from the match algorithm if they fall outside of a specific range. The ranges can be specified independently for the Search Chip and Pattern Chip. While this might not be necessary for a successful registration, excluding pixel data could improve the chances depending on the characteristics of the input images. (i.e., Eliminating the DN range of deep shadow areas in an image).
The ranges are handled via the PVL as follows:
Object = AutoRegistration
Group = PatternChip
Samples = 20
Lines = 20
ValidMinimum = 0.1
ValidMaximum = 0.4
End_Group
Group = SearchChip
Samples = 55
Lines = 55
ValidMinimum = 2.5
ValidMaximum = 10.5
End_Group
End_Group
Valid Pixel Count (ValidPercent)¶
ValidPercent
defines the percentage of valid data that is required
within the Pattern Chip or Search Chip. The parameter can be set in
conjunction with Valid Minimum/Maximum.
The Valid Percent would be handled via the PVL as follows:
Object = AutoRegistration
Group = PatternChip
Samples = 20
Lines = 20
ValidPercent = 80
End_Group
End_Group
Default for ValidPercent
: AutoRegDefaults
Tip
ValidPercent
is specified in the Pattern Chip, as the sub-area
chip of the search area will use the same value.
Fit Chip¶
As the Pattern Chip walks through the sub-regions of the Search Chip , a third chip is generated called a Fit Chip. The Fit Chip will have the same line and sample dimensions as the Search Chip. The Fit Chip is filled with the resulting Goodness of Fit values at every position that the Pattern Chip is walked across the Search Chip computing a fit correlation value.
The highest or lowest Goodness of Fit values within the Fit Chip generally represent the position that best matched between the Pattern and Search areas. It is only good to a maximum of ONE pixel accuracy.
The Fit Chip can be interactively generated and viewed using the image
display application
qnet.
Sub-Pixel Accuracy¶
The attempt to reach a registration of sub-pixel accuracy is performed on the Fit Chip. The Fit Chip is created as the Pattern Chip is walked through Search Chip. The Fit Chip contains the resulting correlation position between the two chips of a maximum of ONE pixel accuracy. In many cases, the actual, ultimate registration may lie somewhere between two pixels. (Refer to: Sub-Pixel Positioning).
The sub-pixel accuracy can be turned off through the PVL settings as follows:
Group = Algorithm
Name = MaximumCorrelation
SubPixelAccuracy = False
End_Group
When the sub-pixel accuracy is turned off, the whole pixel with the best
fit is returned.
Default for SubPixelAccuracy: AutoRegDefaults
Tip
By default, if an ideal Goodness of Fit is found (e.g. Zero (0.0) for Minimum Difference or One (1.0) for Maximum Correlation), we have a perfect fit and assume that it is the best position. In this case, the sub-pixel accuracy phase is omitted.
Surface Modeling¶
A continuous mathematical surface is modeled based on the data within the Fit Chip. A 2nd degree 2-dimensional polynomial is calculated given an NxN window of points extracted from the Fit Chip surrounding the correlation peak (highest value)-which represents the whole pixel registration. An estimate of the true sub-pixel registration position of the chip can be reached based on this surface model.
A perfect 3D linear fit (global maximum) would be represented by a spherical surface with a single high peak.
AutoRegDefaults¶
(As of May 20, 2009)
ValidPercent = 50.0
MinimumZScore = 1.0
Tolerance = Null (no default)
ReductionFactor = 1.0
SubPixelAccuracy = True
SurfaceModelDistanceTolerance = 1.5
SurfaceModelWindows = 5
SurfaceModelEccentricityRatio = 2 (2:1)