- Index
- ImageMagick Examples Preface and Index
- Digital Camera Meta-Data, the EXIF Profile
- Digital Photo Orientation
- Color Improvements
- Photo Conversion Cookbook
One of the prime uses of ImageMagick is the handling and modification of
photographs that were taken with the new modern digital camera.
These cameras generally take quite large, high resolution photos, and include
in them meta-data about the time, scale, zoom, camera, orientation, and so on.
There are even plans to link cameras to mobile phones, so that it can even
make a guess as to where you were when the photo was taken and who might
be in the photo (from what mobile phones are in font of it).
Here we look at the basics of handling digital photos, and even converting
them for other purposes, such as artistic renderings.
Special thanks goes to Walter Dnes, a digital camera user, for his help in the
enhancement of digital photos.
Digital Camera Meta-Data, the EXIF Profile
When a digital camera takes a photo, it also includes a lot of extra
information in the JPEG save file. This meta-data is known as the EXIF
profile, and is provided specifically for photographic labs and development.
The ImageMagick "
identify
" with a "
-verbose
" setting will display
this Exif information.
Here is the EXIF data of a photo I took of a
Pagoda, Kunming Zoo, in Southern China.
identify -format "%[EXIF:*]" pagoda_sm.jpg |\
sed 's/\(.\{46\}\).*/\1/' | column -c 110
|
|
The EXIF data, or any identify output, should be processed in a case
in-sensitive way. Many older versions of IM for example, will output
"EXIF: " (uppercase) rather than "exif: "
(lowercase).
|
Here is a similar example but using 'globbing' (shell-like) expression to
limit output to EXIF fields involving Time...
identify -format "%[exif:*time*]" pagoda_sm.jpg
|
There is a lot of information about this photo in the EXIF profile.
For example
- My camera is a Panasonic ('
Make
'), DMC-LZ1
('Model
')
- The camera was rotated ('
Orientation
').
But I must have corrected that rotation without adjusting the EXIF data.
The camera was also tilted upward slightly, but that info is not recorded.
- The '
FocalLength
' of '37mm
shows that I did not
make use of my cameras 'Optical Zoom' feature. My camera could go up to a
6X optical zoom for a 'FocalLength
' of '366/10
'
or '222mm
'.
- And '
DigitalZoomRatio
' shows I did not digitally zoom either.
- The camera also used a fast 1/8 second '
ExposureTime
', and
an aperture 'MaxApertureValue
' of 3mm, or
'FNumber
' of '5.6
' and a
'ISOSpeedRating
' of '64
'.
- The flash ('
LightSource
') was not used.
- The original image was 1728 by 2304 pixels
('
ExifImageLength
' and 'ExifImageWidth
').
Though the actual image, if you like to check is smaller, so I must have
cropped and/or resized it.
- And probably most importantly, it was taken around 14:05pm on 9th of
July 2005, according to the '
DateTime
' string. That assumes
that I had the cameras time set correctly (which I did).
- More modern cameras may even have a GPS location and posibily a compass
direction of the view!
Also included but not listed above is a small 'thumbnail' preview image that
the camera used on its own display.
There is also features to mark photos you want to be 'developed' or printed by
photographic printers, and to adjust other printing parameters. However this
is rarely used by most people.
Many of these settings can be very useful to users, but the most useful to
people is generally the date and time of the photo. This of course assumes
that the date and time was set correctly on the cemera before the photo was
taken. Also many people are interested in the orientation of the image so it
can be rotated correctly when displayed, and That is what we'll look at next.
All this data, and especially the preview image, can take up quite a lot of
space in the image. And it may be that I don't actually want everyone in the
world knowing that I was in Kunming, China in July 2005. As such you may like
to remove EXIF data from your images before actually publishing it on the
World Wide Web.
Also the size of an image from a digital camera usually very large (and getting
larger), allowing you to print it at photo quality level, but is far too large
for use of the WWW, and especially not for thumbnails. As such unless you want
users to actually be able to print photo quality images, I would not publish
the original image directly.
The above image for example has been cropped and resized for IM examples
usage, but I purposely left the EXIF data intact for the example. Normally
I would strip this information.
Digital Photo Orientation
I have been told that Photoshop will automatically rotate digital images based
on the EXIF '
Orientation
' setting, IM will also do this by
including a "
-auto-orient
" operator, after reading in the image.
However, and this is important
JPEG Format is Lossy
What this means is that any time you decode and save the JPEG file format you
will degrade the image slightly. As a general image processor, IM will always
completely decode and re-encode the format, as such it will always degrade
JPEG images when it re-saves the image. For more information on the nature of
the JPEG format see
JPEG Image File Format.
The point is to only use IM to correct digital photo orientation (using
"
-auto-orient
")
when you are also performing other image modifying operations, such as
Thumbnail Creation,
Annotating Images,
Watermarking or even
Exposure Adjustments.
IM can extract the current orientation (as a number) from the photo using an
Image Property
Escape...
|
identify -format '%[exif:orientation]' pagoda_sm.jpg
| |
|
IM provides a special "
-orient
" operator (use "
-list orientation
" to see possible values).
|
convert pagoda_sm.jpg -orient bottom-right \
-format '%[exif:orientation]' info:
| |
|
These meta-data setting methods, allow you to adjust the orientation of photos
you have modified, especially ones you have rotated. Note that a correctly
orientated photo has an orientation of '
Top-Left
' or 1.
Of course you should not remove the EXIF meta-data (using either "
-strip
" or "
-thumbnail
"), if you plan to
use "
-auto-orient
"
later in the image processing. Use it before stripping the image meta-data.
If you do want to correct the orientation of your photo, without degrading or
otherwise modifying your image, I suggest you use the
JHead program. For example
here I correct a photos orientation, and delete the built-in preview thumbnail
all the digital photos in a directory.
|
The JPEG lossless rotation will only work correctly for images that have
a size that is divisible by 8 or 16. This is true with most (but not all)
digital camera photos. If you try this with an image that is an odd size the
right or bottom edge blocks (containing the partial size) will not be
positioned correctly in the final image, as these block can only exist on
the right or bottom edge.
For an example of this see this specific discussion
|
The
JHead program will
also let you adjust the photos date (if your camera time was set wrong, or you
have travelled to different time zones), extract/remove/replace the preview
thumbnail, set the comment field of the image, remove photoshop profiles, and
do basic image cropping (to remove that stranger exposing himself ;-) so on,
without degrading the JPEG image data.
I recommend this program, or other programs like it (see
Other JPEG Processing Programs), to fix
this information. Just be sure that it does not actually decode/re-encode the
JPEG image data.
One final point about orientation. If you pointed your camera almost straight
up or down, the EXIF orientation setting may not resolve correctly. The same
goes for angled or slanted shots. The orientation (and cameras) just have no
senses for these situations.
Your only choice for such photos is to do the rotates yourself using the lower
level non-lossy "
jpegtrans
", or IM "
-rotate
", and then either reset
the EXIF orientation setting (using
JHead or the IM "
-orient
" operator), or just strip
the EXIF profile.
Other IM Lossy Modifications...
If you are also resizing or otherwise modifying the image, such as reducing
its quality and size for use on the web, then data loss is already a fact.
As such during those operations IM can do similar things, allowing you to do
all the required operations in a single 'load-save' cycle.
Rotate ALL images to landscape -rotate 90\<
portrait -rotate -90\>
Color Improvements
Before proceeding, it is recommended that you first look at
Color Modifications for an introduction to general
color modification techniques that will be used.
Normalizing (using "
-normalize
") high-contrast
line art and graphics can be great. But normalized photos may look unreal,
and, as was said earlier, may not print well either. The "
-contrast-stretch
"
operator can limit the "boundaries" of the normalization, but the "
-levels
" and/or "
-sigmoidal-contrast
"
operator can make "smoother" adjustments (see
Histogram Adjustments for a lower level
discussion of what these operators do).
The above input is courtesy of "Tong" form the IM Mailing List.
Brightening Under-exposed Photos
Contributed by Walter Dnes
Sometimes there simply isn't enough available light to allow for a proper
exposure. At other times, you may have to use shorter exposure times than
optimal, in order to eliminate motion-blur.
Underexposed digital photos can have darker areas preferentially brightened,
without blowing highlights, by using the "
-sigmoidal-contrast
"
operator, with a '
0%
' threshold level. See
Sigmoidal Non-linearity Contrast for more
details.
Here is a minor underexposure example, which was taken at a free concert after
sunset. This has lots of brightly lit areas, which are clear, but also
dark areas I would like to make more visible.
convert night_club_orig.jpg -sigmoidal-contrast 4,0% night_club_fixed.jpg
|
|
As always, you should use a non-lossy format like TIFF or PNG for
intermediate work. The JPEG format is only used here to reduce disk
space and download bandwidth for web publishing.
Select image to see the enlarged version actually used by the examples
rather than the small thumbnail shown.
|
And here is a major underexposed example, which was a night-time shot from my
balcony looking southwards towards the city of Toronto.
convert night_scape_orig.jpg -sigmoidal-contrast 10,0% night_scape_fixed.jpg
|
The main parameter controls the amount of brightening. The more brightening
required the higher value used. And the grainier the output picture will
look. This is due to the smaller pixel errors also being enhanced.
Sigmoidal contrast brightening tends to de-emphasize the red end of the
spectrum. You may end up having to select a parameter that results in the
most natural flesh tones, rather than the brightness level you really want.
In the case of major underexposure, you will end up with a glorified grainy
black-and-white image after brightening. This is a physical limitation of
digital image enhancement. If there's no colour data present, IM won't
generate it for you. In real life the bricks on the right-hand side of my
balcony are reddish, and the trees below are green.
Binning -- Reducing Digital Noise
Contributed by Walter Dnes
A lot of serious photographers are unhappy with the side-effects of
the "megapixel race" by digital camera manufacturers.
Manufacturers pack more megapixels into a digital camera's sensor by making
them smaller. Smaller pixels result in a noisier picture at the same ISO
setting, which forces people to use lower ISO settings. Using lower ISO
ratings to avoid noise requires longer exposure times. This, in turn, means
that most consumer digital cameras are effectively useless indoors
beyond the 10-foot range of their built-in flash for anything except a
still-life picture taken with the camera on a tripod.
Many digital camera users would gladly trade some pixels for less noisy
pictures at higher ISO settings, but the marketeers who control the companies
refuse to consider this as an option.
Fortunately, the trade-off can be done after the fact on digital
photos. The technical term is 'binning'. The simplified theory goes
like so...
- Take an n-by-n grid of pixels, and average their components to obtain
one "super-pixel".
- Signal is proportional to the combined pixel area, which means that
the amount of signal has increased by a factor of n^2
- Noise is random. Which means that it is proportional to the square root
of the combined pixel area, a factor of n. The net result is that SNR
(signal-to-noise ratio) has increased by a factor of n. See Photo
Glossary, Binning for more details.
When a 1600x1200 digital photo is binned down to 800x600 (i.e. a 2x2
grid) the signal-to-noise ratio is doubled. Similarly, a 2560x1920
picture binned 3x3 to 853x640 pixels will have a factor of 3 improvement
in signal-to-noise ratio.
|
In order to make use of binning, the photo image must be a whole number
multiple of the final desired size.
|
In ImageMagick, the special "
-filter
" setting '
box
' will average groups of pixels
down to a single pixel when you "
-resize
" an image (See
Resampling
Filters for details. This means that to do a 'binning' you only need to
resize the image correctly.
Under Construction
Walter Dnes also provided the original script
binn
to perform the calculations, minimally crop the image and perform the
'binning'.
Binning examples 3
Binning examples 4
Photo Conversion Cookbook
Minor Rotation Correction
-- Make a photo more level
Typical situation. You have taken a photo, but the image isn't level, and you
want to correct it.
For example here is a photo I took using a hand held camera in Beijing, 2008,
from the hill in Jingshan Park, immediately behind the Forbidden City. No it
isn't of the Forbidden City itself, but a temple on the other side of the
hill.
Click on the thumbnail, to see a larger image.
Yes the image is small, and you should apply the solution to the
original image not a small thumbnail, but the techniques is the same for any
image. In this case the image needs to be rotated by -1.8 degrees, to correct
it.
Now if you just simply rotate the image you will get a slightly larger image
containing areas of color in the corners, making the correction look obvious
and horible.
convert beijing_tn.png -rotate -1.95 beijing_rotate.png
|
Even if you were to crop the image back to its original size, such as
demonstarted in Simple Image Rotations you
will still get some colored corners.
|
|
The simplist solution would be to now crop that result so as to remove those
borders, but then your image becomes a rather odd size, which is again rather
obvious that something has been done. Though the formula to do that clipping
is not simple, but is demonstrated in
Distortion Rotation Methods.
The better solution is to not only rotate the image, but scale it slightly so
as to produce a rotated image that is the same size as the original.
angle=-1.95
convert beijing_tn.png -distort SRT \
"%[fx:aa=$angle*pi/180;(w*abs(sin(aa))+h*abs(cos(aa)))/min(w,h)], $angle" \
beijing_rot_correction.png
| |
|
And the image is clean looking with a perfectly level wall.
The angle calculation is reasonably straight forward trigonometry, using the
pixel locations at the ends of a long straight line in the image. However
I found that simply rotating the image at various small angles by trial and
error will find a good rotation angle relatively quickly.
When looking how good a particular angle is, take a very close zoomed in look
at the pixels along the line or edge you using. The top of the wall in this
photo. And remember in image rotations a left or anti-clockwise rotation is
negative (due to Y-axis pointing downward).
Also remember that if at all possible, always apply operations to the original
image, avoiding the use intermediate images (and especially intermediate JPEG
images). It is always better to apply any photo modification starting with
the original source than any saved intermediate copy.
Tilt-Shift Effect
-- make scenery look like an artificial model
The 'Tilt-Shift' is a technique which causes an image be be blurred at the
top and bottom, while leaving the center of the image unblurred. It was
originally done in very old bellow type cameras where the lens was tilted
to bring the top and bottom of the image out of focus. Thanks of the
introduction of
Variable Blur Mapping,
added to ImageMagick in v6.5.4-0 this is now easy to do.
If you add to this a very high contrast so as to enhance shadows, and saturate
the colors, a typical result is that a normal image can be made to look
artificial. Almost as if you were taking a photo of a small, highly detailed,
and brightly lit, model.
The first thing we need to do, is enhance the colors in the image to give it a
very high contrast, and perhaps brighten it a bit to make it look like it is
very well lit with strong studio lights.
convert beijing_md.jpg -sigmoidal-contrast 15x30% beijing_contrast.jpg
| |
|
Note how I used a strong
Sigmodial Contrast Operation, to achieve these color effects. I did not
just simply use a linear contrast as I did not want to 'clip' the brightest
and darkest colors of the image. The contrasting value of '
15
'
is a very very strong contrast. I also brightened the image a bit by
offsetting the center of the contrast threshold to a '
30%
' gray
value.
If the colors of the contrast enhanced image does not come out cartoonish
enough, you may like to try increasing the color saturation of the image,
using the
Modulate Operator. This image
did not need it as it as the tiled roof and bright green trees already
provides enough color effects.
If you look at an enlargement of the image (
Click on the thumbnail), you
will see that even just enhancing colors gives the image a feel of artificial
lights, though it does not look like a model, with too much detail to the
cars in the background, and people in the foreground.
Now for the tilt-shift.
For this we prepare a gradient image that is white at the top and bottom, and
black in the middle. Some people might use a linear gradient for this, but
I find a parabolic gradient better.
convert beijing_contrast.jpg \
-sparse-color Barycentric '0,0 black 0,%h white' \
-function polynomial 4,-4,1 beijing_blurmap.jpg
| |
|
Note that I used the original image itself with a two point
Barycentric Sparse Coloring to generate a
linear gradient over the whole image. That linear gradient is then modified
using a basic
Polynomial
Function to make it a parabolic gradient with black in the middle.
Now it is simply a matter of blurring the image according to the blur map to
create a 'tilt-shift' effect. The result is that the original image looks
rather like a scale model, rather than quick snap-shot of the real thing.
convert beijing_contrast.jpg beijing_blurmap.jpg \
-compose Blur -set option:compose:args 10 -composite \
beijing_model.jpg
|
As you can see in the final image, the trees and the buildings look very
artificial, due to the strong colors, while the blurring of the near and far
parts gives the image a 'small' model-like feel to it. Though this must have
been a very detailed model!
The result could have been improved further by performing a
Rotation Correction (see previous) as part of the tilt shift processing.
A perfect camera orientation would simply added to the artificial feel.
Of course you can string all these operations together to to it all in one
command, and avoid temporary files, or loss of quality.
convert beijing_md.jpg -sigmoidal-contrast 15x30% \
\( +clone -sparse-color Barycentric '0,0 black 0,%h gray80' \
-solarize 50% -level 50%,0 \) \
-compose Blur -set option:compose:args 10 -composite \
beijing_model.jpg
|
In the above I replaced the parabolic gradient with a more traditional linear
black-white-gray gradient (with the same slope) to the 'tilt-shift' blur map.
The
Solarize & Level technique was used
to make the linear gradient peak horizontally about 1/3 from the bottom of the
image. However I find that the area of focus in a linear gradient too small
and not very practical.
There are many other way of generating a suitable gradient for a tilt shift
effect. For example using
Resized
Gradients. Or horizontally scaling a
Shepards Sparse Color of single column of pixels. Sine curve gradients
may also be useful.
Speed Optimization
The
Variable Blur Mapping operation is
essentually using a single pass 2-dimentional blurring method (equivelent to
an uniform Gaussaian Blur). However you can get a general speed boost by
doing the bluring operation in two 1-dimensional variable blur operations.
For example here I first blur horizontaly, the vertially...
convert beijing_md.jpg -sigmoidal-contrast 15x30% \
\( +clone -sparse-color Barycentric '0,0 black 0,%h gray80' \
-solarize 50% -level 50%,0 -write mpr:blur_map \) \
-compose Blur -set option:compose:args 10x0 -composite \
mpr:blur_map \
-compose Blur -set option:compose:args 0x10 -composite \
beijing_model_2pass.jpg
|
The result is practially identical (though does differ somewhat), but is a lot
faster to process.
ASIDE: I believe that swaping the operations (blur vertical then horizontally)
will generate a more accurite result for this type of blur mapping. basically
as the horizontal blur is a constant in the direction of that blur pas, so
should be done last.
Problems with Tilt-Shift Effect vs A Real Model
If you examine the resulting photo carefully, you will be able to tell it is
a fake tilt-shift, and not a photo of a real model.
You can see this in that the roof of the larger building is too blurry when
compared to the base of the building. Even though it is about the same
distance as the base. Similarly the base of the 'wall' is more blurry than
the top of the wall. That is it can be seen to be a fake.
The problem is that large vertical objects, should be blurred by the same
amount over the whole surface, and not just variably blurred by height.
Remember the blur gradient is meant to represent the focal depth, or distance
of the various objects in the image, as such the surface of a vertical object
should all be the same 'distance' and thus blurred by the same amount.
To fix I would need to adjust the blur gradient to make those areas have
a with a constant (or near constant) color of the 'base' of that object,
relative to rest of the image. That is vertical surfaces have a constant blur
amount while all the horizontal surfaces have a blur gradient.
Basically the blurred gradient should represent the actual 'depth' of each
point in the image, which for most images is a very complex gradient. This
adjustment can be difficult to achieve, as it most likely requires some human
interpretation of what is a horizontal wall and how far the object is in the
image. It is also unlikely to be easily automated.
What can you do with this effect? Mail me your tilt-shift images! I'll
reference them here. Or perhaps you can correct the tilt-shift faults in the
above example.
PNG-JPEG Layered Images
By separating a large newspaper or magazine page into a text layer that is
saved as a PNG, and an image layer saved as JPG, both using just a white
background, it is possible to use much less disk space than the two images
combined!
More importantally images can use a lossy compression (JPEG), the text
components will remain sharp an clear (PNG).
It sounds silly and weird but it is actually true. The separated images can
save 3 to 4 times the disk space used by a single combined image.
Usually the two images are generated during the publication process as
separate layers. But you can also separate images after the fact too.
The images are just overlayed together...
convert ny_family.jpg ny_family.png -composite ny_family_merged.jpg
|
Select the resulting images to see a larger copy.
This uses a normal
Over Composition, which
requires the PNG (overlay) image to be transparent. This transparency comes
in two forms. Either as a boolean (pure on/off) mask, such as seen in the
above.
Example code for image separation welcome.
Overlapping Photos
-- blurred overlaps of appended photos
Creating a series of overlapping photos (and I don't mean a panarama) is
a common task, especially in web site creation. But is can be tricky to do
unless you have the right knowledge of IM operators.
The simplest method is to use a
Masked
Composite of the two images, and a mask to select which image to overlay.
First however you need to do simple mathematics. For this example, I am using
two thumbnail images 120x90 pixels in size and I want to overlap them
horizontally by 40 pixels. This means the resulting image should be 120 + 120
- 40 pixels wide, or a 200x90 pixel image.
Next we need a mask. This needs to black one one side, white on the other,
with a 40 pixel gradient in the middle, the size of the final output image.
That is 120 pixels - 40 pixel gives an 80 pixel area for each of the two
non-overlapped areas.
So lets generate a masking image...
convert -size 90x80 xc:white xc:black -size 90x40 gradient: \
+swap -append -rotate 90 overlap_mask.png
| |
|
An alternative way of generating the masking image is to use Fred Weinhaus's
"
plmlut
" horizontal gradient generator script. This has finer
controls for the curvature of the gradient rather than a sharp linear gradient
I generate above.
Now that all of the math is out of the way, all that is left is to do a three
image masked composition, using the mask we just generated. However we will
also need to enlarge the destination (left) image so as to provide enough
space for the overlapping right image (any color), and position the second
image correctly using the appropriate gravity (right, or '
East
').
convert holocaust_tn.gif -extent 200x90 spiral_stairs_tn.gif \
overlap_mask.png -gravity East -composite overlap_photos.jpg
|
And we now have two images, which are overlapped using a linear gradient.
Of course the two commands can be merged into a single command, so that you
don't need to save the 'mask' intermediate image. This is left as an exercise
to the reader.
A slight improvement is to use a more curved gradient over a larger overlap
between the images. This reduces the sharp change visible at the start and
end of the overlap area of the final image. Especially with images contain
large areas of very different colors.
For example, this uses some
Distorted
Gradient techniques to not only generate a smoother gradient curve,
but also to rotate that gradient so as to have a highly angled overlap.
convert -page +0-15 -size 1x30 gradient: \
-sigmoidal-contrast 5,50% -contrast-stretch 0 \
-set option:distort:viewport 180x90-90-45 \
+distort SRT 115 +repage \
holocaust_tn.gif -extent 180x90 +swap \
spiral_stairs_tn.gif +swap \
-gravity East -composite overlap_angled.jpg
| |
|
Yes, the above is rather complex, but it shows just what is possible.
If you plan to do more than two images, a better method is to use the mask to
directly set the transparency of the second and later images. The multiple
images can then be overlaid together using a techniques seen in
Layered Image Examples.
Some of these techniques do not require you calculate the final image size, as
IM can do this for you. You only need to make sure you position the images
correctly.
For example, here I add a 30 pixel gradient to a second and third image,
requiring the images to be placed every 90 pixels (width 120 minus 30 pixel
overlap) from each other. When all images are given the appropriate
transparency and positioning, we just
Mosaic
the layers together (all offsets are positive), letting IM figure out the
final canvas size.
convert -size 90x90 xc:white -size 90x30 gradient: -append -rotate 90 \
hatching_tn.gif \
\( chinese_chess_tn.gif -clone 0 \
-compose CopyOpacity +matte -composite -repage +90+0 \) \
\( holocaust_tn.gif -clone 0 \
-compose CopyOpacity +matte -composite -repage +180+0 \) \
\( spiral_stairs_tn.gif -clone 0 \
-compose CopyOpacity +matte -composite -repage +270+0 \) \
-delete 0 -compose Over -mosaic overlap_series.jpg
|
Rather than pre-calculating the positions of the overlapping masked images,
you can use techniques found in
Append
Overlap, as well as
Incrementally
Calculated Positions for longer image sequences.
Final Notes:
Overlapping photos like this works best for images with a reasonably common
overall color.
Also you may notice that for the images at either end of the sequence,
a centered subject may not look very centered due to the overlap on one side
of the image only. This problem can be improved either by fading the outside
edge of those images into transparency, or chopping of some of outside edge to
help re-center the subject of those images.
ASIDE: It may be that doing the composition in a different colorspace
may work better. Anyone like to experiment and report on your results, good
or bad?
Double Exposures
-- mixing multiple photos of the same scene
With old time film based cameras, there was a technique where a picture was
take two or more times without 'rolling' the film. This allowed you to create
what was known as double exposures, where two images taken at slightly
different times were merged together. The result was often a ghosting or
dimming of parts of the image which moved or changed.
However with careful control of the subjects in the image, the lighting
effects, and even the development process, it became possible to make some
very weird or even 'impossible' photos. With digital images it is even
easier as you have even better control of the images.
Basically...
Seeing may be believing, but cameras lie!
For example suppose I wanted an image in which I appear in twice! Well that is
easy to do. Here for example are the thumbnails of two quick photos I took
specifically for this example, using a tripod and timer, which I'll use
directly.
Perhaps you can supply a better more amusing photo set?
I will apply the double exposure techniques directly to these thumbnails,
though more typically I would do this using original image files as inputs, so
as to get a result of the highest quality.
Now if I used a traditional film-like 'double exposure' with an old style
camera, the result would be an average of these two images, generating
see-thru 'ghosts' of myself. Here is the digital simulation of this
technique...
convert anthony_1.jpg anthony_2.jpg -average anthony_ghosts.jpg
| |
|
However, what if I don't want ghosts, but properly solid images of myself.
Well then you need to use a mask to select which parts you want to come from
which image.
This mask can be generated in two ways. You can just manually create the mask
by dividing the image along the static or unchanging parts. A rather simple
matter in this particular case...
convert -size 100x90 xc: -draw 'rectangle 0,0 50,89' \
-blur 0x3 anthony_mask.jpg
| |
|
Note that I blurred the mask, so as to 'feather' the cut-over between the two
images. And here I use a
Masked Composition
to merge the images.
convert anthony_1.jpg anthony_2.jpg anthony_mask.jpg \
-composite anthony_doubled.jpg
| |
|
How if you had two (or more) family photos, where some people had eyes closed,
were speaking, pulling faces, or just looking away. You could pick and choose
each 'head' from different images and merge the multiple images to form a
montage, so as to get a photo where everyone is looking at the camera, and
have their eyes open.
By swapping the input images, or just negating the mask, you can remove me
completely from the image, so get an unrestricted view of the static
background.
convert anthony_2.jpg anthony_1.jpg anthony_mask.jpg \
-composite anthony_removed.jpg
| |
|
This can be handy when taking photos of a public monument, where you can't
afford the expense of crowd control. Just take lots and lots of photos from a
tripod, and hopefully you can combine them to remove everyone from the scene!
An alturnative to generating a background image which you have hundredes of
images (video) is to just create an average of all the images. This Turn all
the people or other transient objects into a light haze of 'ghosts'. That can
be an interesting effect in and of itself, but not always what is wanted.
An average image can be an useful step, as once you have it you can compare it
against each individual image to mask out the person (transient object) from
each frame, before again combining the backgrounds together, to create a clean
(no haze) background image).
A major discussion on automatically generating a 'clean background' from video
images is in IM Discussion Forum on
Creating a Referance Image and
Extracting change events.
With a clean background photo, we we can threshold a difference image to mask
out the parts of the image that changed. You may need to use some further
blurring and threshold to expand that mask appropriately to cover not only the
object within the image, but any shadows or reflections it may cast on the
background scenery. A little trial and error may also be needed to get it
right.
convert anthony_removed.jpg anthony_2.jpg \
-compose difference -composite \
-threshold 5% -blur 0x3 -threshold 20% -blur 0x3 \
anthony_automask.jpg
| |
|
Now lets use this mask to mix my 'ghosts' image with the original image so it
looks like my conscience is 'haunting' me for making such 'impossible'
pictures.
convert anthony_1.jpg anthony_ghosts.jpg anthony_mask.jpg \
-composite anthony_haunted.jpg
| |
|
As a final point, all the above techniques assumes the photos were taken from
a camera that was locked down securely on a stationary tripod. If this was
not the case, but just taken from a hand held position, I can guarantee that
the images will not match-up or 'align' properly, no matter how hard you tried
to do it. In such cases you may require some
Affine or even
Perspective
distortion of at least one of the two images to get the backgrounds to align
properly. The more complex the background, the more exacting the needed
re-alignment.
If a flash was used, or the day was cloudy with variable light, you may also
need some brightness adjustments to the photo. The cause is that most cameras
'auto-adjust' the brightness of the images, and a flash, or variable light can
change its handling of the 'auto-level' adjustment for each and every image.
As a final example, here is another image I created from two separate photos,
of my nephew fencing with himself, in front of a climbing wall. As I was
holding the camera and used a flash, I did need to do some affine distortion
adjustments, as well as slight brightness adjustment to get the seamless
result you see.
Jacob vs Jacob
If you were trying to decide if this photo was fake or not, you would look at
the lighting, shadows and reflections. In the above, a close examination of
the floor will show that the right 'Jacob' does not have a proper reflection
on the floor (it was clipped by the photos edge). But you would really need to
study the photo well to notice this!
Now think of the possibilities you can use this 'double exposure' technique
for. For example how about some
Funny
Mirrors. Email me your results!
If you like to get into this further the research paper "
Interactive Digital Photomontage", goes into using "Double Exposures" (or
as it terms it "photo montage"), but making use of user selections expanded
using "image segmentation", to select what parts of the image is to come from
where.
One example is if you have a number of photos of a large group of people, in
each photo someone does not 'look good'. You can use this technique to
select which person comes from which image so that you can get a perfect group
photo where everyone is: facing front, with eyes open, and smiling!
Protect Someone's Anonymity
-- fuzzing out some part of a photo
The above technique of using a 3 image composite mask can also be used in
other ways. For example you can 'pixelate' and image, then use a mask to
limit the effect to just the face of a person, so as to "Protect their
Identity".
convert zelda_tn.gif -scale 25% -scale 400% zelda_pixelate.gif
convert zelda_tn.gif -gamma 0 -fill white \
-draw 'circle 65,53 50,40' zelda_face_mask.gif
convert zelda_tn.gif zelda_pixelate.gif zelda_face_mask.gif \
-composite zelda_anonymity.png
|
Of course you can do this all in one go, and even smooth the change from
pixelated to normal. For example..
convert zelda_tn.gif \( +clone -scale 25% -scale 400% \) \
\( +clone -gamma 0 -fill white \
-draw 'circle 65,53 50,40' -blur 10x4 \) \
-composite zelda_anonymity.jpg
| |
|
Of course rather than pixelate the offending part, you can also blur the area
instead. Just replace the two "
-scale
" operators with a single "
-blur
" to fuzz out the details.
This technique replacing a masked area can also be used to remove unwanted
text and logos from images. For details see
Hole Filling.
Add a Texture to an Image
The
Hardlight alpha compositing method or
even any of the various
Lighting Composition
Methods provide ways to give an image a texture pattern.
For example here I add a texture of course fabric to a photo I took of a
pagoda at the Kunming Zoo, in southern China.
convert tile_fabric.gif -colorspace gray -normalize \
-fill gray50 +level 35% texture_fabric.gif
composite texture_fabric.gif pagoda_sm.jpg \
-tile -compose Hardlight photo_texture.jpg
|
Note that if you want to actually
tile the texture over the image you
need to use the "
composite
" command rather than the more
versatile "
convert
" command, though there are a number of other
ways to
Tile Images in Memory using
convert.
Also note that when adding a texture like this, the smaller details in the
original photo can be lost by excess noise of the overlaid texture, textures
should ge hept either simple, or their effect appropriatally moderated, such
as the
Decontrasting Level Adjustment
used above.
To use an image pattern as a texture it should be modified so that a perfect
gray color is used for areas that is unchanged in the original image. That is
the average color of the image should be about 50% gray. In the example I
demonstrate one way that you can do this with just about any tileable image,
though this specific method may not always work well.
Such textures can be found all over the web, as various background patterns
for web pages. They may not even look like a texture, be colorful, or even
very bright or very dark. After adjustment however you will find that you can
get some very interesting effects.
Just as we did previously, you can limit what parts of an image is actually
textured by creating an appropriate mask. For example lets create a mask of
just the near 'white' sky in the pagoda photo.
convert pagoda_sm.jpg -fuzz 10% -transparent white \
-alpha extract -negate pagoda_mask.png
convert pagoda_sm.jpg photo_texture.jpg pagoda_mask.png \
-composite photo_texture_masked.jpg
|
Now imagine an picture of a lady wearing a dress. You can get any pattern
shade it appropriately, and then overlay that on the original image so as to
replace the dress with a completely different design.
Of course there are lots of variations on the above to achieve the final
result, and which specific technique you use is up to you, but the basic idea
is the same. Texture the image, mask and overlay the result.
As an aside, I also recommend you look at the
Overlay alpha composing method, which is simply the same as
Hard_Light composition, but with the two
images swapped. There is also a lot of other
Shading Composition Methods that can be used to texture an image in
various ways.
Chroma Key Masking
-- Modifying by areas of specific color
The photo to the left was given by an user in a
IM Forum Discussion. he wanted to change
the color of the girls shirt, which was a nice 'pink' color. The problem is the
color is not just 'pink' but a whole range of different shades of 'pink'.
As you have seen above, to make changed to an image, the first step is
typically generating an appropriate mask of the area you are interested in.
Here I will use a technique known as
Chroma Key to generate
mask that specific color.
This technique generally looks for a specific color in an image for use as the
mask. It is also the technique used for 'blue' and 'green' screen effects
used extensively on TV and in Movies.
This basically involves extracting the 'Hue' by
Separating Channel Images, then looking
up the 'hue shade' wanted. For example...
convert shirt.jpg -colorspace HSL -channel Hue -separate shirt_hue.jpg
|
However this Hue image has a couple of problems.
- First a 'pink' color is very close to 'red' which is at the division where
Hue 'rolls over'. To ensure this is not a problem I use Modulate to adjust the hue away from
that 'discontinuity' in the hue.
This is not a problem for extracting a 'chroma key' for 'green' or 'blue'
screens.
- This 'pink' color is also not a highly saturated color, but has a very low
saturation value. This means its 'hue' is not as strong as it should be.
- The other problem is the gray background!!!!! Gray is has very little hue,
so I need to remove any areas with little to no saturation from my final
mask, or I'll be changing things in the background.
Note that this is technically not needed if I limit changes to hue rolls,
which does not effect unsaturated colors.
In short, the input image would have worked better with a brighter stronger
color that was also not as simial to skin (or hair) color. A strong blue or
green shirt for example. But I will work with what I was given.
|
|
So lest extract and combine the two channel masks. Note that Hue = Gray64
after the image hues was 'rolled' using module, and Saturated = Black for the
grey background.
convert shirt.jpg -modulate 100,100,33.3 -colorspace HSL \
-channel Hue,Saturation -separate +channel \
\( -clone 0 -background none -fuzz 5% +transparent grey64 \) \
\( -clone 1 -background none -fuzz 10% -transparent black \) \
-delete 0,1 -alpha extract -compose multiply -composite \
shirt_mask.png
|
That just leaves a number of small isolated 'specks' that can be removed with
some Morphology Smoothing
(-morphology Smooth Square ). It isn't perfect but it will do the
job. The better way would be to edit the mask by hand to clean it up.
Now a mask can be used with Composite Masking
much like we did with Double Exposures and Anonymity examples above.
However If you are using a mask to modify an existing image (without
distorting, or changing the images size), then it is easier to use it to
define what areas are un-writable. These are known as Clip or Write Masks (see "-mask "
|
|
Here I cleanup the previous mask of the small defects (optional), and negate
it to define what areas I want to 'write protect'. Then I set this mask,
shift the hues to turn 'pink' into a 'light blue' color, and save the
resulting image.
convert shirt_mask.png -morphology Smooth Square \
-negate shirt_write_mask.png
convert shirt.jpg -mask shirt_write_mask.png \
-modulate 100,100,25 +mask shirt_blue.jpg
|
Yes there is a slight 'pink' border, especially in the inside sleeve. Also
a small area of skin on her arm turned a rather dark blue. Basically these
are mask defects, and with a little more work in perfecting the mask you can
fix these problems. But it is not bad result.
One method of generating a better mask is to use a much larger higher
resolution image. When the resulting image is later resized these small
defects will (hopefully) also be reduced to insignificance.
|
|
The real problem with this specific example, is the 'key color' is so close to
a normal skin color you are really just asking for trouble! This is why
people using this technique use 'green' and 'blue' screens, as those colors
are as different as possible from 'skin' color of people in front of the
screen.
Note that you are better off NOT using JPEG as your source or working images.
Really JPEG should only be used for your final images only! This is part of
the reason why so many 'mask defects' was generated in the first place.
Green Screen
Future example, using Chroma Key Masking of a 'green screen background'.
Expanded from the wikipedia artical, Chroma Key
Real problems in 'green screen' handling is the 'color spill', with fine light
color hair (blonde) and semi-transparent areas producing the worse color spill
effects.
Simplistic Colorspill removal (color fix)
g(r,g,b) => (r, min(g, b), b)
Alpha determination...
a(r,b,g) => K0 * b − K1 * g + K2
Using values of 1.0 for all K coefficients is good initial guess.
As the Background color is well known, and once the 'alpha' is known
you can use techniques shown in Background Removal using Two Backgrounds to remove any 'green screen
halo' that may be present better that the first color formula.
Artist Charcoal Sketch of Image
The
Charcoal Sketch Transform, offers
users a very simple way of generating a simplified gray-scale rendering of the
image.
It does not work well for 'busy images' but for simpler images it can produce
a very striking result.
convert holocaust_sm.jpg -charcoal 5 charcoal.gif
|
Children's Color-In Outline Image
In a long discussion about
Generating
Coloring-In Pages on the IM Users Forum, the following cookbook recipe was
developed to convert a simple photo into something children can color in.
Here is the best result we have so far, applied to a photo I took of the
holocaust memorial, Berlin.
convert holocaust_sm.jpg \
-edge 1 -negate -normalize \
-colorspace Gray -blur 0x.5 -contrast-stretch 0x50% \
color-in.gif
# For heavily shaded pictures...
# #-segment 1x1 +dither -colors 2 -edge 1 -negate -normalize \
|
The final operations in the above attempt to smooth out the lines and improve
the overall result.
Of course the above technique is only useful for images with good sharp color
changes, and preferably a higher resolution image than I used above.
For cartoon images that already have black outlines with a light colored
background, the use of
Edge Detection with
the above method will directly produce a 'twinning' effect of the black
outlines. You can see this effect in the twinned lines of tiles on the path
leading into the memorial, in the lower-left corner.
This is an artifact of the way
Edge
Detection works, and you can see more examples of this in that section of
IM Examples.
The solution is to negate images of this type before using "
-edge
" to outline the colored
areas.
convert piglet.gif -background white -flatten \
-colorspace Gray -negate -edge 1 -negate -normalize \
-threshold 50% -despeckle \
-blur 0x.5 -contrast-stretch 0x50% \
color-in_cartoon.gif
|
I also "
-threshold
"
so I can then remove individual dots that "
-edge
" seem to like to generate.
After that I again attempt to smooth out the aliased lines in the image.
The above was added to in a discussion on
GIMP Photocopy Filter to make use of the
Compose Divide method, to find outlines.
convert taj_mahal_sm.png -colorspace gray \
\( +clone -blur 0x2 \) +swap -compose divide -composite \
-linear-stretch 5%x0% photocopy.png
|
The "
-linear-stretch
" operation in the above adjusts how black the
dark areas of the images will be, while the "
-blur
" 'sigma' defines the shading
sharpness.
Pencil Sketch
Using a
Photoshop (PSP)
tutorial on converting images to
Pencil Sketches,
dognose from the
IM Users Forum, managed to create the
equivalent ImageMagick commands. Here is his conversion, simplified into a
few IM commands, allowing you to batch process lots of images into a 'artists
pencil sketch' form.
First we need a special "
pencil.gif
" image. This can take a long
time, so for this example I made it a bit smaller, while preserving its
ability to be tiled across larger images. See
Modifying Tile Images for details of the techniques.
This only needs to be done once and can then be re-used. As such you can
generate a much larger one for your own use, so as to avoid any tiling
effects. Ideally make it as large as the images you plan to convert.
convert -size 256x256 xc: +noise Random -virtual-pixel tile \
-motion-blur 0x20+135 -charcoal 1 -resize 50% pencil_tile.gif
| |
|
Now it is only a matter of overlaying and blending this 'pencil' shading image
with a photo. The pencil image is tiled to make a canvas the same size as the
image we are processing. Then it is applied to the image using techniques
found in
Tiled Canvases. This is then merged
into a gray-scaled copy of the original image.
convert pagoda_sm.jpg -colorspace gray \
\( +clone -tile pencil_tile.gif -draw "color 0,0 reset" \
+clone +swap -compose color_dodge -composite \) \
-fx 'u*.2+v*.8' sketch.gif
|
Note that as the "
-blend
"
operator of the "
composite
"
command is not available to the "
convert
" command, I opted to do
the equivalent using the DIY "
-fx
" operator. There are probably better, faster but more
complicated ways of doing this. (suggestions are welcome)
This is not the final version, as the operator misses some edge enhancement
aspects needed for outline some of the more lighter but sharp color changes in
the image. Can you improve the above?
The above algorithm was built into IM as an artistic transform "
-sketch
", though without the
"
-resize
" smoothing for
the generated 'pencil tile'...
convert pagoda_sm.jpg -colorspace gray -sketch 0x20+120 sketch_new.gif
|
Vignette Removal
When taking photos (digital or otherwise, the camera lens generally darkens
the edges and corners of the image. This is called 'vignetting'. In fact this
lens effect is so common, it is often faked on purpose using the "
-vignette
" operator. See the
Vignette Transform.
Martin Herrmann <Martin-Herrmann@gmx.de> wanted to remove camera
vignetting from the photos. Basically he took a photo of a white sheet of
paper in a bright light without using a flash. He then wanted to combine this
with his actual photos to brighten the edges and corners of the image
appropriately.
Basically what we want to do is divide the original photo by the grey-scale
image of the photo of the brightly lit white piece of paper and it will then
brighten the parts of the image by the amount that the 'white paper' photo was
darkened.
This is basically the compose method '
Divide
' which divides the 'source' image by the 'background'
image. For example,
However as the photo of the 'white paper' will probably not be a true white,
and you probably do not want to brighten the image by this 'off-white' color.
To fix this we need to multiply the divisor image by its the center
pixel color.
Here is the final solution provided to Martin, which used the very slow
FX DIY Operator. This pre-dated the addition of
a
Divise Compose Method which can be used to
speed up this process enormously.
The white photo was also grey scaled to remove any color distortion as well,
note that I changed the ordering which will also preserve any 'meta-data' that
was in the original (as it is the 'destination' image in this case.
convert vegas_orig.jpg \( nikon18-70dx_18mm_f3.5.jpg -colorspace Gray \) \
-fx '(u/v)*v.p{w/2,h/2}' vegas_fixed_fx.jpg
| |
|
If you look carefully at the enlarged photos, particularly the top-left and
top-right 'sky' corners, you can see the vignetting effects, and the
correction that was made.
It is not a perfect solution, and could use a little more tweaking. For
example rather than using a scaling pixel, we could pre-process the 'white
page' image, and also adjust it for a better vignette removal result.
Note that using JPEG is not recommended for any sort of photographic work, as
the format can introduce some artifacts and inconsistencies in the results.
The format is only good for storage and display of the final results.
A major discussion on correcting vignettation is in the IM User Forums in the
discussion
Algorithmic vignetting
correction for pinhole cameras?.
Things that can effect vignettation include...
- Distance of film from lens, further away means more light spread.
- Area of the aperture 'circle' (lens or pinhole) due to angle of light.
- Arrangement of camera material around the aperture. For example
the lens holder or pinhole thickness.