snibgo's ImageMagick pages

Blending pyramids

Images can be blended by making Laplacian pyramids, blending these to make a third, and collapsing it.

Feathering, which uses a gradient to modulate alpha in order to blend images, is a quick but crude technique that blends data of all frequencies by the same spatial amount. It either cuts large-scale features into distracting seams, or creates ghosting of small-scale features. We can reduce both these problems by blending Laplacian pyramids.

The technique is also called multi-resolution splining.

This page builds on material in:

Scripts on this page assume that the version of ImageMagick in %IMDEV% has been built with various process modules. See Process modules.

References:

Sample inputs

Take two images of the same size.

set SRC1=toes.png
toes.png
set SRC2=bp_src2.png

%IM%convert ^
  toes.png ^
  -negate -flip -flop ^
  %SRC2%
bp_src2.pngjpg

Suppose we want to blend the left half of image A with the right half of image B.

Hard cut:

%IM%convert ^
  %SRC1% ^
  %SRC2% ^
  -crop 2x1@ +repage ^
  -delete 1-2 ^
  +append ^
  bp_hard.png
bp_hard.pngjpg

The cut is sharp. It is the sharpest feature in the image. It is distracting.

Alpha blend:

%IM%convert ^
  %SRC1% ^
  %SRC2% ^
  ( -clone 0 ^
    -sparse-color bilinear ^
      "0,0,Black %%[fx:w-1],0,White" ^
    +write mpr:GRAD ^
  ) ^
  -alpha off ^
  -composite ^
  -size 10x10 xc:White ^
  ( mpr:GRAD ^
    -crop x30+0+0 +repage ) ^
  -append ^
  bp_alph1.png
bp_alph1.pngjpg
%IM%convert ^
  %SRC1% ^
  %SRC2% ^
  ( -clone 0 ^
    -sparse-color bilinear ^
      "0,0,Black %%[fx:w-1],0,White" ^
    -level 30%%,70%% ^
    +write mpr:GRAD ^
  ) ^
  -alpha off ^
  -composite ^
  -size 10x10 xc:White ^
  ( mpr:GRAD ^
    -crop x30+0+0 +repage ) ^
  -append ^
  bp_alph2.png
bp_alph2.pngjpg
%IM%convert ^
  %SRC1% ^
  %SRC2% ^
  ( -clone 0 ^
    -sparse-color bilinear ^
      "0,0,Black %%[fx:w-1],0,White" ^
    -level 45%%,55%% ^
    +write mpr:GRAD ^
  ) ^
  -alpha off ^
  -composite ^
  -size 10x10 xc:White ^
  ( mpr:GRAD ^
    -crop x30+0+0 +repage ) ^
  -append ^
  bp_alph3.png
bp_alph3.pngjpg

The images are superimposed across the width of the blend. This causes ghosting, where we see detail from one image bleeding through to the other.

Blend halves of opaque images

From each image, we make a Laplacian pyramid. Then we make a third pyramid, where each grid in this new pyramid is the left half of the corresponding grid in the first pyramid, appended to the right half of the grid in the second pyramid. Then we collapse this third pyramid.

set pyPREFIX=bp_blr_

call %PICTBAT%mkLapPyr %SRC1% bp_pyr1.tiff
call %PICTBAT%mkLapPyr %SRC2% bp_pyr2.tiff

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IM%convert ^
    bp_pyr1.tiff[%%i] ^
    bp_pyr2.tiff[%%i] ^
    -crop 2x1@ +repage ^
    -delete 1-2 ^
    +append ^
    -strip ^
    blr_out_%%i.tiff

  set FILE_LIST=!FILE_LIST! blr_out_%%i.tiff
)

%IM%convert %FILE_LIST% bp_pyr3.tiff

sed -e 's/bp_pyr2/bp_pyr3/g' blr_mklp_recon.scr >blr_recon.scr

%IM%convert @blr_recon.scr toes_inv.png
toes_inv.png

If we want a diagonal blend, or any other shape, we can make a black and white mask.

set spCORNERS=^
0,0,Black,^
%%[fx:w-1],0,gray(50%%),^
0,%%[fx:h-1],gray(50%%),^
%%[fx:w-1],%%[fx:h-1],White

%IM%convert ^
  %SRC1% ^
  -sparse-color bilinear ^
    "%spCORNERS%" ^
  bp_diag_msk.png
bp_diag_msk.png

As before, we can use this mask for an alpha blend:

%IM%convert ^
  %SRC1% ^
  %SRC2% ^
  bp_diag_msk.png ^
  -alpha off ^
  -composite ^
  bp_diag1.png
bp_diag1.pngjpg
%IM%convert ^
  %SRC1% ^
  %SRC2% ^
  ( bp_diag_msk.png ^
    -level 45%%,55%% ^
  ) ^
  -alpha off ^
  -composite ^
  bp_diag2.png
bp_diag2.pngjpg
%IM%convert ^
  %SRC1% ^
  %SRC2% ^
  ( bp_diag_msk.png ^
    -threshold 50%% ^
  ) ^
  -alpha off ^
  -composite ^
  bp_diag3.png
bp_diag3.pngjpg

Threshold the mask:

%IM%convert ^
  bp_diag_msk.png ^
  -threshold 50%% ^
  bp_diag_msk_th.png
bp_diag_msk_th.png

We can blend the two Laplacian pyramids using this thresholded mask. Of course, the mask is larger than most grids in the pyramid so we can't use it directly. Instead, we make a Gaussian pyramid and use each Gaussian grid as a mask for compositing the Laplacian grids into a third Laplacian, which we then collapse to reconstruct an image.

call %PICTBAT%mkGausPyr bp_diag_msk_th.png bp_msk.tiff

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IM%convert ^
    bp_pyr1.tiff[%%i] ^
    bp_pyr2.tiff[%%i] ^
    bp_msk.tiff[%%i] ^
    -alpha off -composite ^
    blr_out_d_%%i.tiff

  set FILE_LIST=!FILE_LIST! blr_out_d_%%i.tiff
)

%IM%convert %FILE_LIST% bp_pyr3.tiff

%IM%convert @blr_recon.scr bp_diag_inv.png
bp_diag_inv.pngjpg

The result is a smooth blend with no ghosting.

We can do the same trick with a Gaussian of the un-thresholded mask, but the result is the same as the alpha blend of the two opaque images using the gradient mask, without any use of pyramids:

call %PICTBAT%mkGausPyr bp_diag_msk.png bp_msk.tiff

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IM%convert ^
    bp_pyr1.tiff[%%i] ^
    bp_pyr2.tiff[%%i] ^
    bp_msk.tiff[%%i] ^
    -alpha off -composite ^
    blr_out_d_%%i.tiff

  set FILE_LIST=!FILE_LIST! blr_out_d_%%i.tiff
)

%IM%convert %FILE_LIST% bp_pyr3.tiff

%IM%convert @blr_recon.scr bp_diag_inv2.png
bp_diag_inv2.pngjpg

As usual, the scripts above could be modified to create a command or script that runs convert just once overall instead of once per octave.

Minimum error

Instead of using a Gaussian pyramid generated from a mask, we can do a minimum-error boundary cut (MEBC) of each pair of grids of the Laplacian pyramids. See Dark paths: boundary cut.

Here is the simple MEBC of an ordinary pair of images (not pyramids):

%IMDEV%convert ^
  %SRC1% ^
  %SRC2% ^
  ( -clone 0-1 ^
    -compose Difference -composite ^
    -process 'darkestmeander m 0 a' ^
    +write bp_dp.png ^
    -fill White ^
    -draw 'color 0,0 floodfill' ^
    -negate ^
    +write bp_dp_msk.png ^
  ) ^
  -compose Over -composite ^
  bp_me1.png
bp_dp.png bp_dp_msk.png bp_me1.pngjpg

The boundary is obvious. I find this jagged boundary more distracting than the straight version, even if the two sides of the boundary are a better match.

Here we apply MEBC to each grid of the two Laplacian pyramids, creating a third, which we collapse. Each grid has its own MEBC.

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IMDEV%convert ^
    bp_pyr1.tiff[%%i] ^
    bp_pyr2.tiff[%%i] ^
    ^( -clone 0-1 ^
       -compose Difference -composite ^
       -process 'darkestmeander m 0 a' ^
       -fill White ^
       -draw 'color 0,0 floodfill' ^
       -negate ^
    ^) ^
    -alpha off ^
    -compose Over -composite ^
    -depth 16 ^
    blr_out_mebc_%%i.tiff

  set FILE_LIST=!FILE_LIST! blr_out_mebc_%%i.tiff
)

%IMDEV%convert %FILE_LIST% bp_pyr3.tiff

%IM%convert @blr_recon.scr bp_mebc.png
bp_mebc.pngjpg

The effect is virtually seamless, with less ghosting than the equivalent alpha-blend version. For some images, ghosting is virtually eliminated. For example:

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IMDEV%convert ^
    bp_pyr1.tiff[%%i] ^
    bp_pyr2.tiff[%%i] ^
    ^( -clone 0-1 ^
       -compose Difference -composite ^
       -process 'darkestmeander m 0 a' ^
       -fill White ^
       -draw 'color 0,0 floodfill' ^
    ^) ^
    -alpha off ^
    -compose Over -composite ^
    -depth 16 ^
    blr_out_mebc_%%i.tiff

  set FILE_LIST=!FILE_LIST! blr_out_mebc_%%i.tiff
)

%IMDEV%convert %FILE_LIST% bp_pyr3.tiff

%IM%convert @blr_recon.scr bp_mebc2.png
bp_mebc2.pngjpg

There is very slight visible ghosting (a nearly horizontal boundary near the centre).

Blend Over

Suppose we have these two images:

set WALL=dpr_darker_wall.png
dpr_darker_wall.pngjpg
%IM%convert ^
  dpt_lvs_sm.jpg ^
  dps_lvs_msk.png ^
  -alpha off ^
  -compose CopyOpacity -composite ^
  -background Black -alpha background ^
  bp_lvs.png

set LVS=bp_lvs.png
bp_lvs.pngjpg

We can composite the leaves over the wall. First in the obvious way, then by corresponding pyramid layers with the script proc2LapPyr.bat. The fourth parameter to the script must be a valid IM process that takes two images and returns one image. Here, we use "-composite".

Throughout the following, we use mkpKEEP_ALPHA_RES. See Laplacian pyramids with transparency.

set mkpKEEP_ALPHA_RES=1
%IM%convert ^
  %WALL% ^
  %LVS% ^
  -composite ^
  bp_wall_lvs.png
bp_wall_lvs.pngjpg
call %PICTBAT%proc2LapPyr ^
  %WALL% ^
  %LVS% ^
  . "-composite" ^
  bp_lpo.png

Low frequencies (colours) have mingled between the images.
The images have, in a sense, "flattened" together.

bp_lpo.pngjpg

With a suitable mask, we can transition the effect.

Make a mask for the transition.

call %PICTBAT%featherEdge ^
  dps_lvs_msk.png 75 75 bp_fe.png

%IM%convert ^
  bp_fe.png ^
  -solarize 50%% ^
  -evaluate Multiply 2 ^
  -sigmoidal-contrast 5,50%% ^
  bp_fe.png
bp_fe.png

Apply the transition mask.

%IM%convert ^
  bp_wall_lvs.png ^
  bp_lpo.png ^
  bp_fe.png ^
  -composite ^
  bp_lpo2.png
bp_lpo2.pngjpg

The colour flattening occurs in both directions, so the wall becomes somewhat leaf-coloured and the leaves become somewhat wall-coloured. Can we hold one constant, so only the other changes? The wall is fully opaque; the leaves are somewhat transparent, in all levels of the pyramid. We can change the alpha channel by any process that doesn't change alpha=0 or alpha=100%, so fully transparent leaf pixels and fully opaque leaf and wall pixels remain as they are.

call %PICTBAT%proc2LapPyr ^
  %WALL% ^
  %LVS% ^
  . ^
  "-channel A -evaluate Pow 0.1 +channel -composite" ^
  bp_lpo3.png

Power <1 increases all alpha values except alpha=0 and alpha=100%.
The leaves are almost unchanged; the wall has changed to match the leaves.

bp_lpo3.pngjpg
call %PICTBAT%proc2LapPyr ^
  %WALL% ^
  %LVS% ^
  . ^
  "-channel A -evaluate Pow 10 +channel -composite" ^
  bp_lpo4.png

Power >1 decreases all alpha values except alpha=0 and alpha=100%.
The wall is almost unchanged; the leaves have changed to match the wall.

bp_lpo4.pngjpg

As before, a transition mask could be used to control the placement of the effect.

In these last few examples, I have repeatedly used proc2LapPyr.bat on the same input images, which repeatedly makes the same pyramids. Applications that need multiple processing of pyramids could be more economical.

Matching gain and bias

Compositing Laplacian layers will tend to colour-match the result. We can do this explicitly, for example by the Gain and bias technique. Adjusting the brightness and contrast (gain and bias) of the leaves to match the wall, or the wall to match the leaves, looks like this:

call %PICTBAT%imgGainBias ^
  dpt_lvs_sm.jpg ^
  %WALL% ^
  bp_gb1.png
bp_gb1.pngjpg
call %PICTBAT%imgGainBias ^
  %WALL% ^
  dpt_lvs_sm.jpg ^
  bp_gb2.png
bp_gb2.pngjpg

Alternatively, we can build Laplacian pyramids from the images, adjust the gain and bias of the grids, and collapse the result. We do this with a script matchLapGainBias.bat.

call %PICTBAT%matchLapGainBias ^
  dpt_lvs_sm.jpg ^
  %WALL% ^
  . ^
  bp_lgb1.png
bp_lgb1.pngjpg
call %PICTBAT%matchLapGainBias ^
  %WALL% ^
  dpt_lvs_sm.jpg ^
  . ^
  bp_lgb2.png
bp_lgb2.pngjpg

Adjusting the gain and bias via pyramids (instead of with the simple images) has slightly decreased contrast in the first result, and increased contrast in the second. It exagerates the effect slightly.

Matching histograms

Another technique for colour-matching is shown in Process modules: matching histograms. Simple histogram-matching of the leaves to match the wall, or the wall to match the leaves, looks like this:

call %PICTBAT%matchHisto ^
  dpt_lvs_sm.jpg ^
  %WALL% ^
  bp_mh1.png
bp_mh1.pngjpg
call %PICTBAT%matchHisto ^
  %WALL% ^
  dpt_lvs_sm.jpg ^
  bp_mh2.png
bp_mh2.pngjpg

Compared to the gain-and-bias technique, both results have higher contrast.

Simpler examples

The effects can perhaps be seen more clearly with simpler examples. We create a 300x300 image for a background, and another image 100x100 extended with transparency to 300x300 for a foreground. We composite the foreground over the background in the conventional way, then via Laplacian pyramids. We show both the outputs. To some outputs, we overlay a blue square showing the original boundary of the foreground.

Throughout, we use:

set mkpKEEP_ALPHA_RES=1

set BOUNDARY=-stroke Blue -fill None -draw "rectangle 100,100 199,199"

set pyGRAPHIC=1
%IM%convert ^
  -size 300x300 xc:gray(50%%) ^
  bp_s1a.png

%IM%convert ^
  -size 100x100 xc:#f00 ^
  -gravity Center ^
  -background None -extent 300x300 ^
  bp_s1b.png

%IM%convert ^
  bp_s1a.png bp_s1b.png ^
  -composite ^
  bp_s1c.png

call %PICTBAT%proc2LapPyr ^
  bp_s1a.png bp_s1b.png ^
  . "-composite" ^
  bp_s1d.png

The result is an almost uniform colour.

bp_s1c.pngjpg bp_s1d.pngjpg
%IM%convert ^
  -size 100x100 gradient:#f00-#fff ^
  -gravity Center ^
  -background None -extent 300x300 ^
  bp_s2b.png

%IM%convert ^
  bp_s1a.png bp_s2b.png ^
  -composite ^
  bp_s2c.png

call %PICTBAT%proc2LapPyr ^
  bp_s1a.png bp_s2b.png ^
  . "-composite" ^
  bp_s2d.png

%IM%convert ^
  bp_s2d.png ^
  %BOUNDARY% ^
  bp_s2d.png
bp_s2c.pngjpg bp_s2d.pngjpg
%IM%convert ^
  -size 300x300 gradient:white-black ^
  bp_s3a.png

%IM%convert ^
  bp_s3a.png bp_s1b.png ^
  -composite ^
  bp_s3c.png

call %PICTBAT%proc2LapPyr ^
  bp_s3a.png bp_s1b.png ^
  . "-composite" ^
  bp_s3d.png
bp_s3c.pngjpg bp_s3d.pngjpg
%IM%convert ^
  -size 300x300 gradient:white-black ^
  bp_s4a.png

%IM%convert ^
  -size 100x100 gradient:#f00-#fff ^
  -region 50x50+50+0 -negate +region ^
  -gravity Center ^
  -background None -extent 300x300 ^
  bp_s4b.png

%IM%convert ^
  bp_s4a.png bp_s4b.png ^
  -composite ^
  bp_s4c.png

call %PICTBAT%proc2LapPyr ^
  bp_s4a.png bp_s4b.png ^
  . "-composite" ^
  bp_s4d.png

%IM%convert ^
  bp_s4d.png ^
  %BOUNDARY% ^
  bp_s4d.png
bp_s4c.pngjpg bp_s4d.pngjpg
%IM%convert ^
  -size 100x100 gradient:#f00-#fff ^
  -region 50x50+50+0 -negate +region ^
  -gravity Center ^
  -background None -extent 300x300 ^
  bp_s5b.png

%IM%convert ^
  bp_s1a.png bp_s5b.png ^
  -composite ^
  bp_s5c.png

call %PICTBAT%proc2LapPyr ^
  bp_s1a.png bp_s5b.png ^
  . "-composite" ^
  bp_s5d.png

%IM%convert ^
  bp_s5d.png ^
  %BOUNDARY% ^
  bp_s5d.png
bp_s5c.pngjpg bp_s5d.pngjpg
%IM%convert ^
  -size 100x100 gradient:#f00-#fff ^
  -function sinusoid 1.5,90,0.5,0.5 -negate ^
  -gravity Center ^
  -background None -extent 300x300 ^
  bp_s6b.png

%IM%convert ^
  bp_s1a.png bp_s6b.png ^
  -composite ^
  bp_s6c.png

call %PICTBAT%proc2LapPyr ^
  bp_s1a.png bp_s6b.png ^
  . "-composite" ^
  bp_s6d.png

%IM%convert ^
  bp_s6d.png ^
  %BOUNDARY% ^
  bp_s6d.png
bp_s6c.pngjpg bp_s6d.pngjpg
%IM%convert ^
  -size 300x300 gradient:white-black ^
  -function sinusoid 4.5,90,0.5,0.5 -negate ^
  bp_s7a.png

%IM%convert ^
  bp_s7a.png bp_s6b.png ^
  -composite ^
  bp_s7c.png

call %PICTBAT%proc2LapPyr ^
  bp_s7a.png bp_s6b.png ^
  . "-composite" ^
  bp_s7d.png

%IM%convert ^
  bp_s7d.png ^
  %BOUNDARY% ^
  bp_s7d.png
bp_s7c.pngjpg bp_s7d.pngjpg
%IM%convert ^
  -size 300x300 gradient:white-black -rotate 90 ^
  -function sinusoid 4.5,90,0.5,0.5 -negate ^
  bp_s8a.png

%IM%convert ^
  bp_s8a.png bp_s6b.png ^
  -composite ^
  bp_s8c.png

call %PICTBAT%proc2LapPyr ^
  bp_s8a.png bp_s6b.png ^
  . "-composite" ^
  bp_s8d.png

%IM%convert ^
  bp_s8d.png ^
  %BOUNDARY% ^
  bp_s8d.png
bp_s8c.pngjpg bp_s8d.pngjpg

As previous, but with centre
100x100 of background transparent.

%IM%convert ^
  -size 300x300 gradient:white-black -rotate 90 ^
  -function sinusoid 4.5,90,0.5,0.5 -negate ^
  -alpha set ^
  -region 100x100+100+100 ^
    -alpha transparent ^
  +region ^
  bp_s9a.png

%IM%convert ^
  bp_s9a.png bp_s6b.png ^
  -composite ^
  bp_s9c.png

call %PICTBAT%proc2LapPyr ^
  bp_s9a.png bp_s6b.png ^
  . "-composite" ^
  bp_s9d.png

%IM%convert ^
  bp_s9d.png ^
  %BOUNDARY% ^
  bp_s9d.png

Foreground detail has spread further into the background;
background detail has spread less far into foreground.

bp_s9c.pngjpg bp_s9d.pngjpg

Detail at the lowest frequency (solid colour) spreads entirely between images. Detail at the highest frequency doesn't spread at all. Detail at intermediate frequencies spreads by intermediate amounts.

Where the foreground was red and white, why has the white become cyan? Because the lowest frequency of the background is the mean colour, which is 50% gray. This has spread entirely into the centre 100x100 pixels, mixing with the mean from the foreground, resulting in an overall mean of roughly 50% gray. We can show this:

%IM%convert bp_s6d.png -crop 96x97+102+102 +repage -scale "1x1^!" txt: 
# ImageMagick pixel enumeration: 1,1,65535,srgba
0,0: (35330,32861,32861,65535)  #8A02805D805DFFFF  srgba(54%,50%,50%,1)

But the higher frequencies have not spread as much. The higher frequencies of the foreground contained two areas with "more red than average" and two with "less red than average". The average has changed from a light red to roughly 50% gray, and "less red than 50% gray" is cyan.

As with the leaves on the wall, we can adjust the alpha.

%IM%convert ^
  bp_s8a.png bp_s6b.png ^
  -channel A -evaluate Pow 0.1 +channel -composite ^
  bp_a1c.png

call %PICTBAT%proc2LapPyr ^
  bp_s8a.png bp_s6b.png ^
  . "-channel A -evaluate Pow 0.1 +channel -composite" ^
  bp_a1d.png

%IM%convert ^
  bp_a1d.png ^
  %BOUNDARY% ^
  bp_a1d.png

With power <1, foreground detail spreads further into the background;
background detail spreads less far into the foreground.

bp_a1c.pngjpg bp_a1d.pngjpg
%IM%convert ^
  bp_s8a.png bp_s6b.png ^
  -channel A -evaluate Pow 10 +channel -composite ^
  bp_a2c.png

call %PICTBAT%proc2LapPyr ^
  bp_s8a.png bp_s6b.png ^
  . "-channel A -evaluate Pow 10 +channel -composite" ^
  bp_a2d.png

%IM%convert ^
  bp_a2d.png ^
  %BOUNDARY% ^
  bp_a2d.png

With power >1, foreground detail spreads less far into the background;
background detail spreads further into the foreground.

bp_a2c.pngjpg bp_a2d.pngjpg

With ImageMagick, we can compose an image with transparency over an equal-sized image, or we can use an equal-sized third image (typically grayscale) as a mask that determines transparency. We have a similar choice with pyramids: the second Laplacian pyramid might have transparency, or it can be opaque and we use a third pyramid as a mask. This third pyramid will be Gaussian. We do this with the script proc3LapPyr.bat that takes three images, builds the pyramids, applies a process level-by-level, and collapses the resulting pyramid. For the process, we use "-alpha off -composite".

Corresponding to examples above, we make three opaque 300x300 images. The third (bp_llg1m.png) is the mask. As above, we show a result made from a conventional IM compose, then the result made with pyramids (with a blue square to mark the boundary).

%IM%convert ^
  -size 300x300 gradient:white-black -rotate 90 ^
  -function sinusoid 4.5,90,0.5,0.5 -negate ^
  bp_llg1a.png

%IM%convert ^
  -size 300x300 gradient:#f00-#fff ^
  -function sinusoid 4.5,90,0.5,0.5 -negate ^
  bp_llg1b.png

%IM%convert ^
  -size 300x300 xc:Black ^
  -fill White -draw "rectangle 100,100 199,199" ^
  bp_llg1m.png

%IM%convert ^
  bp_llg1a.png bp_llg1b.png bp_llg1m.png ^
  -alpha off -composite ^
  bp_llg1c.png

call %PICTBAT%proc3LapPyr ^
  bp_llg1a.png bp_llg1b.png bp_llg1m.png ^
  . "-alpha off -composite" ^
  bp_llg1d.png

%IM%convert ^
  bp_llg1d.png ^
  %BOUNDARY% ^
  bp_llg1d.png
bp_llg1c.pngjpg bp_llg1d.pngjpg

When using a mask, we adjust the spread by adjusting the mask.

%IM%convert ^
  bp_llg1a.png bp_llg1b.png bp_llg1m.png ^
  ( +clone -evaluate Pow 0.1 ) ^
  +swap +delete ^
  -alpha off -composite ^
  bp_llg2c.png

call %PICTBAT%proc3LapPyr ^
  bp_llg1a.png bp_llg1b.png bp_llg1m.png ^
  . ^
  "^( +clone -evaluate Pow 0.1 ^) +swap +delete -alpha off -composite" ^
  bp_llg2d.png

%IM%convert ^
  bp_llg2d.png ^
  %BOUNDARY% ^
  bp_llg2d.png
bp_llg2c.pngjpg bp_llg2d.pngjpg

As always, operations can be performed in L*a*b* space. Blah.

Cleanup

set mkpKEEP_ALPHA_RES=

Conclusion

Blending Laplacian pyramids is a fast and effective method for seamlessly combining photographs, especially when used with minimum error boundary cut.

The heaviest processing work is in creating the pyramids, and darkestmeander is also expensive. All the image processing on this web page was performed in about ten seconds on my laptop.

Scripts

For convenience, .bat scripts are also available in a single zip file. See Zipped BAT files.

matchLapGainBias.bat

rem Makes version of %1 that resembles %2,
rem by gain-and-bias method.
rem %3 is optional output filename.

@if "%2"=="" findstr /B "rem @rem" %~f0 & exit /B 1

@setlocal enabledelayedexpansion

rem @call echoOffSave

call %PICTBAT%setInOut %1 mlgb

set INFILE2=%2

if not "%3"=="" set OUTFILE=%3

set PYR_PARAM=%~3
if "%PYR_PARAM%"=="." set PYR_PARAM=
if "%PYR_PARAM%"=="" set PYR_PARAM=. . . .

if not "%4"=="" set OUTFILE=%4

set TMPDIR=%TEMP%
rem for /F "usebackq" %%L in (`cygpath -u %TMPDIR%`) do set CYGTMPDIR=%%L

set pyPREFIX=mlgb_
call %PICTBAT%mkLapPyr %INFILE2% %pyPREFIX%pyr2.tiff %PYR_PARAM%
call %PICTBAT%mkLapPyr %INFILE% %pyPREFIX%pyr1.tiff %PYR_PARAM%

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  call %PICTBAT%imgGainBias ^
    %pyPREFIX%pyr1.tiff[%%i] ^
    %pyPREFIX%pyr2.tiff[%%i] ^
    %TMPDIR%\%pyPREFIX%out_%%i.tiff

  if ERRORLEVEL 1 exit /B 1

  set FILE_LIST=!FILE_LIST! %TMPDIR%\%pyPREFIX%out_%%i.tiff
)

%IM%convert %FILE_LIST% %pyPREFIX%pyr1.tiff

%IM%convert @%pyPREFIX%mklp_recon.scr %OUTFILE%


call echoRestore

endlocal & set mlgbOUTFILE=%OUTFILE%

matchLapHisto.bat

rem Makes version of %1 that resembles %2,
rem by histogram-matching grids in Laplacian pyramids with four parameters %3.
rem %4 is optional output filename.

@if "%2"=="" findstr /B "rem @rem" %~f0 & exit /B 1

@setlocal enabledelayedexpansion

rem @call echoOffSave

call %PICTBAT%setInOut %1 mlh

set INFILE2=%2

if not "%3"=="" set OUTFILE=%3

set PYR_PARAM=%~3
if "%PYR_PARAM%"=="." set PYR_PARAM=
if "%PYR_PARAM%"=="" set PYR_PARAM=. . . .

if not "%4"=="" set OUTFILE=%4

set TMPDIR=%TEMP%
for /F "usebackq" %%L in (`cygpath -u %TMPDIR%`) do set CYGTMPDIR=%%L

set pyPREFIX=mlh_
call %PICTBAT%mkLapPyr %INFILE2% %pyPREFIX%pyr2.tiff %PYR_PARAM%
call %PICTBAT%mkLapPyr %INFILE% %pyPREFIX%pyr1.tiff %PYR_PARAM%

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IMDEV%convert ^
    %pyPREFIX%pyr1.tiff[%%i] ^
    ^( -clone 0 ^
       -process 'mkhisto cumul norm' ^
    ^) ^
    ^( %pyPREFIX%pyr2.tiff[%%i] ^
       -process 'mkhisto cumul norm' ^
       -process 'mkhisto cumul norm' ^
    ^) ^
    ^( -clone 1-2 -clut ^) ^
    -delete 1-2 ^
    -clut ^
    %CYGTMPDIR%\%pyPREFIX%out_%%i.tiff

  if ERRORLEVEL 1 exit /B 1

  set FILE_LIST=!FILE_LIST! %TMPDIR%\%pyPREFIX%out_%%i.tiff
)

%IM%convert %FILE_LIST% %pyPREFIX%pyr1.tiff

%IM%convert @%pyPREFIX%mklp_recon.scr %OUTFILE%


call echoRestore

endlocal & set mlhOUTFILE=%OUTFILE%

procLapPyr.bat

rem From image %1, makes Laplacian pyramid with four parameters %2,
rem applies IM processing %3,
rem collapses pyramid into output %4.

@if "%1"=="" findstr /B "rem @rem" %~f0 & exit /B 1

@setlocal

@call echoOffSave

call %PICTBAT%setInOut %1 plp


set PYR_PARAM=%~2
if "%PYR_PARAM%"=="." set PYR_PARAM=
if "%PYR_PARAM%"=="" set PYR_PARAM=. . . .

set IM_PROC=%~3
if "%IM_PROC%"=="." set IM_PROC=
if "%IM_PROC%"=="" set IM_PROC=+write info:

if not "%4"=="" set OUTFILE=%4

set TMPDIR=%TEMP%

set pyPREFIX=plp_
call %PICTBAT%mkLapPyr %INFILE% %pyPREFIX%pyr.tiff %PYR_PARAM%

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IM%convert ^
    %pyPREFIX%pyr.tiff[%%i] ^
    %IM_PROC% ^
    %TMPDIR%\%pyPREFIX%out_%%i.tiff

  set FILE_LIST=!FILE_LIST! %TMPDIR%\%pyPREFIX%out_%%i.tiff
)

%IM%convert %FILE_LIST% %pyPREFIX%pyr.tiff

%IM%convert @%pyPREFIX%mklp_recon.scr %OUTFILE%


call echoRestore

endlocal & set plpOUTFILE=%OUTFILE%

proc2LapPyr.bat

rem From images %1 and %2, makes Laplacian pyramid with four parameters %3,
rem applies IM processing %4 (which should create one image from two),
rem collapses pyramid into output %5.
@rem
@rem Updated:
@rem   15 May 2016 use %IML% for @script.

@if "%2"=="" findstr /B "rem @rem" %~f0 & exit /B 1

@setlocal

@call echoOffSave

call %PICTBAT%setInOut %1 p2lp

set INFILE2=%2

set PYR_PARAM=%~3
if "%PYR_PARAM%"=="." set PYR_PARAM=
if "%PYR_PARAM%"=="" set PYR_PARAM=. . . .

set IM_PROC=%~4
if "%IM_PROC%"=="." set IM_PROC=
if "%IM_PROC%"=="" set IM_PROC=+write info:

if not "%5"=="" set OUTFILE=%5

set TMPDIR=%TEMP%

set pyPREFIX=p2lp_
call %PICTBAT%mkLapPyr %INFILE2% %pyPREFIX%pyr2.tiff %PYR_PARAM%
call %PICTBAT%mkLapPyr %INFILE% %pyPREFIX%pyr1.tiff %PYR_PARAM%

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IM%convert ^
    %pyPREFIX%pyr1.tiff[%%i] ^
    %pyPREFIX%pyr2.tiff[%%i] ^
    %IM_PROC% ^
    %TMPDIR%\%pyPREFIX%out_%%i.tiff

  set FILE_LIST=!FILE_LIST! %TMPDIR%\%pyPREFIX%out_%%i.tiff
)

%IM%convert %FILE_LIST% %pyPREFIX%pyr1.tiff

%IML%convert @%pyPREFIX%mklp_recon.scr %OUTFILE%


call echoRestore

endlocal & set p2lpOUTFILE=%OUTFILE%

proc3LapPyr.bat

rem From images %1 and %2 and mask %3,
rem   makes Laplacian pyramid of first two
rem   and Gaussian pyramid of third, with four parameters %4,
rem applies IM processing %5 (which should create one image from two),
rem collapses pyramid into output %6.
@rem
@rem Updated:
@rem   15 May 2016 use %IML% for @script.

@if "%3"=="" findstr /B "rem @rem" %~f0 & exit /B 1

@setlocal

rem @call echoOffSave

call %PICTBAT%setInOut %1 p3lp

set INFILE2=%2
set MSKFILE=%3

set PYR_PARAM=%~4
if "%PYR_PARAM%"=="." set PYR_PARAM=
if "%PYR_PARAM%"=="" set PYR_PARAM=. . . .

set IM_PROC=%~5
if "%IM_PROC%"=="." set IM_PROC=
if "%IM_PROC%"=="" set IM_PROC=+write info:

if not "%6"=="" set OUTFILE=%6

set TMPDIR=%TEMP%

set pyPREFIX=p3lp_
call %PICTBAT%mkGausPyr %MSKFILE% %pyPREFIX%msk.tiff %PYR_PARAM%
call %PICTBAT%mkLapPyr %INFILE2% %pyPREFIX%pyr2.tiff %PYR_PARAM%
call %PICTBAT%mkLapPyr %INFILE% %pyPREFIX%pyr1.tiff %PYR_PARAM%

set FILE_LIST=

for /L %%i in (0,1,%mklpNUM_OCTAVES%) do (

  %IM%convert ^
    %pyPREFIX%pyr1.tiff[%%i] ^
    %pyPREFIX%pyr2.tiff[%%i] ^
    %pyPREFIX%msk.tiff[%%i] ^
    %IM_PROC% ^
    %TMPDIR%\%pyPREFIX%out_%%i.tiff

  set FILE_LIST=!FILE_LIST! %TMPDIR%\%pyPREFIX%out_%%i.tiff
)

%IM%convert %FILE_LIST% %pyPREFIX%pyr1.tiff

%IML%convert @%pyPREFIX%mklp_recon.scr %OUTFILE%


call echoRestore

endlocal & set p3lpOUTFILE=%OUTFILE%

All images on this page were created by the commands shown, using:

%IM%identify -version
Version: ImageMagick 6.9.5-3 Q16 x86 2016-07-22 http://www.imagemagick.org
Copyright: Copyright (C) 1999-2015 ImageMagick Studio LLC
License: http://www.imagemagick.org/script/license.php
Visual C++: 180040629
Features: Cipher DPC Modules OpenMP 
Delegates (built-in): bzlib cairo flif freetype jng jp2 jpeg lcms lqr openexr pangocairo png ps rsvg tiff webp xml zlib

Source file for this web page is blendpyr.h1. To re-create this web page, execute "procH1 blendpyr".


This page, including the images, is my copyright. Anyone is permitted to use or adapt any of the code, scripts or images for any purpose, including commercial use.

Anyone is permitted to re-publish this page, but only for non-commercial use.

Anyone is permitted to link to this page, including for commercial use.


Page version v1.0 21-Oct-2014.

Page created 04-Aug-2016 18:10:52.

Copyright © 2016 Alan Gibson.