coreg

This program co-regsisters two images using an image-wide averaged sample/line translation with the TRANSLATE option or a set of variable sample/line translations with the WARP option. The program computes local translations spaced evenly throughout the image. The number and spacing of local translations is user defined. This allows for many output options including 1) directly creating the translated image, 2) creating a control network which can be used in other programs (e.g., qnet, warp), especially if the translation is not constant across the image, or 3) creating a flat-field file usable in spreadsheets or plotting packages to visualize magnitude and direction of varying translations. For the TRANSLATE option, this implies that the internal geometry of both images be nearly the same so that the translation can be computed. That is, this program will not work if the translation varies significantly throughout the image. If condition of near constant translation is met then the translation can be computed to sub-pixel accuracy. If the internal geometry of both images do not match well, or a simple line/sample shift is not sufficient to register the images, the WARP option is a better choice than TRANSLATE because this uses the local translations to perform a rubber-sheet transformation. This option must be used with caution.

NOTE: This program can utilize many different techniques for computing the translation. It is recommended that you review the Pattern Matching page from the "Related Objects and Documents" section below. It is essential for using this application to understand how to create a "registration definition" file and how to size your search and pattern chips. We will continue with the discussion of functionality of this program assuming the reader has a fundamental knowledge of Automatic Registration. Below we give an example of a defintion file (DEFFILE):

      Object = AutoRegistration
        Group = Algorithm
          Name         = MaximumCorrelation
          Tolerance    = 0.7
        EndGroup

        Group = PatternChip
          Samples = 20
          Lines   = 20
          ValidMinimum = 1400
          ValidPercent = 75
        EndGroup

        Group = SearchChip
          Samples = 90
          Lines   = 90
          ValidMinimum = 1400
          ValidPercent = 75
        EndGroup
      EndObject
    

For further discussion of the parameters used in this DEFFILE, see the Pattern Matching document. Briefly, the example DEFFILE will allow a successful registration only where the MaximumCorrelation algorithm's goodness-of-fit result is >=0.7, pixel value is >1400, and at least 75% of the pixels in both the PatternChip or the SearchChip are valid pixels.

This program requires two input cubes, one which will be translated (FROM) and the other is held fixed (MATCH). The images must have the same number of samples and lines and can have only one band (use cube attributes if necessary). A grid will be defined across the held image using either the user parameters, ROWS and COLUMNS, or calculated based on the image size and the search chip size as follows: COLUMNS = (image samples - 1) / search chip samples + 1, and similarly for ROWS. Conceptually, the sparse grid defined by ROWS and COLUMNS will be laid on top of both images with even spacing between the rows (or columns) and but no row will touch the top or bottom of the image. That is, the grid is internal to the image.

At each grid intersection, the local translation will be computed. This is done by centering the search chip at the grid intersection for the image to be translated (FROM) and centering the pattern chip at the grid intersection for the held image (MATCH). The pattern chip is walked through the search chip to find the best registration (if any). Again, see the Pattern Matching document for further details. The local translation is recorded at all grid intersections that had a successful registration. The results are written to a control network and/or flat-file if requested. The average of the local translations is then used to compute an overall sub-pixel translation which can be applied to the FROM image and written as the output image (TO).

Some tips to improve odds of a successful registration are provided. In general a small pattern chip size makes registration more difficult. Depending on your dataset, 20x20 is probably a good starting point. The larger the translation, the larger the search chip size will need to be; if your translation is only a couple of pixels, you should make the search chip only slightly larger than the pattern (e.g., 25x25 vs 20x20). However if the translation is large you will need to expand the search area. For example, if the translation is roughly 45 pixels and your pattern is 20x20 the search area should be, roughly, 20+2*45 or 110x110.

This example shows the coreg application. The rows and columns parameters are left as default.

Sample,Line,TranslatedSample,TranslatedLine,SampleDifference,LineDifference,GoodnessOfFit
211.875,164.283,200,133,-11.8751,-31.2829,0.991597
211.766,429.437,200,398,-11.7661,-31.4369,0.995987
        

The above text file is the ffile.txt file created when coreg is ran. The flat file is comma seperated and can easily be imported into excel.

coreg from=./lunar1.cub match=./lunar2.cub t=out.cub flatfile=ffile.txt

First Input image	First Input image for coreg Parameter Name: FROM This is the 800 by 800 input image to be translated for the coreg example.
Second Input image	Second Input image for coreg Parameter Name: MATCH This is the 800 by 800 input image to be held for the coreg example.

Output image showing results of the coreg application.

Output image for coreg Parameter Name: TO This is the 800 by 800 output image that results.