Examples

Example 1: Aligning a Pair of Images

We will align examples/images/z00.tif to examples/images/z01.tif:

   

First, visually compare the two 1024x1024 pixel images using:

cd examples
display images/z00.tif images/z01.tif

or your favorite image viewer.  (Hitting the space bar in display will alternate between the two images.)  Although similar, these adjacent sections differ slightly in content and have a noticeable difference in rotation.  We will use the following workflow for aligning these two images.

z00 will be our source image and z01 our target, so the first step is find an approximate rotation, scale, and translation using find_rst:

echo "z00 z01 z00_z01" >pairs00_01.lst
mkdir -p cmaps
find_rst -pairs pairs00_01.lst -images images/ -output cmaps/ -max_res 1024 \
         -summary cmaps/summary.out

This will yield cmaps/z00_z01.map  (cmaps is short for "coarse maps").  The max_res option indicates the resolution at which matching will take place, and the processing time is strongly dependent on it.  Higher resolutions than 1024 willl take much longer, and need only be used if the images lack large-scale features, but have sufficient small-scale features in common.  A summary is written to cmaps/summary.out.  Generally, if the quality is >0.1, then the find_rst program has been successful at finding a match.  For certain datasets a quality of 0.05-0.1 is still acceptable, but for others may indicate a failure to find a good match.  Usually, a quality of < 0.05 indicates that no good match was found.

The next step is to obtain a non-linear registration of the two images.  This is done by running register, giving it as an initial map the one found by find_rst:

mkdir -p maps
register -pairs pairs00_01.lst -images images/ -output maps/ -initial_map cmaps/ \
         -distortion 2.0 -output_level 6 -depth 6 -quality 0.5 \
         -summary maps/summary.out

This produces maps/z00_z01.map.  The distortion option is a penalty weight to give to distorted maps, so higher values will tend to produce more rigid and uniform maps.  The output_level of 6 indicates that final map points should be space every 2^6 = 64 pixels in x and y.  The depth of 6 determines what image resolution is used to a given map resolution.  Here, 6 (output_level) - 6 (depth) = 0, so the full resolution images will be used when computing the final maps.  (Coarser maps are internally generated prior to the final map since register uses a multiresolution algorithm)  Again, a summary of the image pairs processed is written to maps/summary.out.  A correlation score indicating success varies widely from dataset to dataset, but generally values above 0.5 are good.  Sometimes values as low as 0.1 can still be good -- this just depends on the image characteristics, high-frequency noise being one contributor to lowered correlation scores.

After running register, it is good to inspect the pairwise registration maps using the inspector program:

inspector -pairs pairs00_01.lst -input images/ -map maps/ -tif -no_gray

The left mouse button will alternate between the original z00.tif and a warped z01.tif that has been transformed according to maps/00_01.map to match up with z00.tif.  The correspondence cannot be exact because the images are slightly different in content, but the overall image features should match up well visually, which they do here.  ESC will exit the program.

To generate a warped version of z00.tif that is in registry with z01.tif, we use apply_map:

ln -s z00_z01.map maps/z00.map
apply_map -image z00 -images images/ -maps maps/ -output warped/ -region 1024x1024+0+0

apply_map, when it tries to process image z00, will look in images/ for z00.tif and then in maps/ for z00.map.  We want it to use the map from z00 onto z01, and thus we first make a symlink from z00.map to z00_z01.map.  The warped output image is placed in warped/z00.tif.  The -region option is necessary so that the output is cropped to a 1024x1024 image corresponding to the size and position of z01.tif.  If we had not included that option, the output would have been 1041x1064 in size and offset by 3 pixels in x and 24 in y.  To view the warped z00 and compare it with the original z01, we can use the following command:

display warped/z00.tif images/z01.tif

Hitting the space bar will advance from one to the next.  The black margins are areas where there were no corresponding pixels in z00.  The images displayed should look similar to: