ImageMagick Notes

ImageMagick is often weird but never stupid. Its idiosyncratic command line processing style is subtle but reasonable once understood. The command (pretty much all commands from the Imagemagick family) is followed by arguments. These do not follow Unix/Gnu tools conventions. There are image specifiers. These either load or write an image or use (synthesize) a built-in one. They usually can infer the correct image type if you’re not being complicated. There are simple processing operators that work individually on each image that has been loaded thus far into the image stack. This happens immediately after the operator is read which why it is important to get your order right and have images to be processed loaded.

Here is an example command.

This reads in the (pseudo) files logo: and wizard:. Then comes the geometry operator which operate immediately on all images in the stack — in this case, on each of them. The final argument is the output but what will it do with two processed images? It turns out that it creates two files called out-1.png and out-2.png.

There are some operators that take the entire stack of (perhaps) many images and do something to the collection that produces just one output file. These include -composite and -combine naturally enough but many others too.

The image stack itself can be managed with some special operator arguments like -delete, -clone, -reverse, -swap, etc.

Some operators are settings that are applied to subsequent operators. A good example is -quality.

Here are some images that can be used with ImageMagick that it generates on the fly for testing or programmatic applications. They add new images to memory and do not modify existing ones without further action. To see what some of the generative images look like try something like this.

There are lots of patterns too which can be used with something like: display -size 99x99 pattern:bricks

ImageMagick is very handy for finding out properties of image files. Often viewers scale images up or down and it’s hard to tell what you really have. The identify program cuts through the mystery and gives you the solid truth about an image.

Or if you need more detailed info.

If you need something very fancy look into the -format specifier.

Sometimes you need to know if some known thing like a watermark or logo is present on the image. Here is how you can extract the value of a single pixel at 50,1000.

Note that txt: is a special output file that changes behavior to dump output.

I have a bunch of images which should contain blue (0,0,255) masks. For some reason some of them are washed out to (32,32,248). This will find all such pixels and replace them with proper pure blue.

Another problem with those images is that sometimes while making the masks, Gimp used a fuzzy tool — so instead of getting crisp edge, I get a bunch of pixels with softening alpha values.

You can see a histogram of how many colors your image set has like this.

To normalize it so that they all have only 2 colors, use this.

Another fun problem I ran into is that when converting a bunch of 2 color masks, anytime the entire image was only one color (and perhaps specifically white), the image converted automatically to grayscale. This caused problems and needed to be cured. Here are some possible helpful hints though this problem seems very hard.

To get a sense of what the type codes are look at the "Properties:" section of identify -verbose. Note that the star after the +profile operator is (confusingly) to remove all profiles and is not related to the files. Profiles show up in exif data and can be put there by messy programs. This is hopefully a way to get rid of it to standardize your images.

One of the most important executables in the ImageMagick suite is called convert and as you might expect one of its primary functions is to convert images from one type to another.

What if you have an annoying EPS (Encapsulated Postscript) "image" that you’d like to convert to a real bitmap. If the sizes come out all wrong, consider setting the -density flag. Make sure it comes first!

Perhaps the most common application for ImageMagick that I have is resizing images. These are general techniques:

I wish I knew how exactly the -geometry specification worked but if you use something like 400x400 then the maximum dimension, in X or Y, will be 400 with the other scaled to preserve the aspect ratio.

The official manual says, "Geometry is a very special option. The operator behaves slightly differently in every IM command, and often in special and magical ways." Great.

Another way to resize images that could be useful if you knew what percent of the original size you wanted.

There is also -scale which seems to be a less thoughtful version of resize. Even more lightweight in trying to keep visual fidelity is the -sample operator. The -resample operator is for changing resolution or density (there is also a -density setting).

If you need thumbnails, there is a built-in -thumbnail operator that is designed to make that easy.

There is a -magnify operator that tries to round the corners of magnified pixels and maybe other tricks needed for making images have more pixels without adding any new real information.

Sometimes you get an image and you know that for whatever reason, it has been scaled on the X axis by -1. To put that right use the -flop option of convert, no option argument needed. The matching -flip works on the vertical axis but that seems a lot less useful to me in real life. Maybe try both if your camera is upside-down.

The official web help is helpful. Cropping is a bit weird and like many other aspects of ImageMagick, there is much about it I do not understand.

The case where cropping most naturally comes in handy is when trying to isolate a section of an image. Perhaps you have a series of video frames and you only want the part that is looking out a car window above the hood or something. To do this you can start by interactively determining the region of interest. Use xwininfo and then click the display (maybe feh or eog or display) and it will tell you a lot including the "window ID". Now you use xev -id 0x3a00005. Then when you mouse around the image in the display program, you’ll get coordinate display of the pointer. Once you’ve figured out just what you need, you can use something like this.

This will extract a 250x125 part of the window and if the window is 1024x768, it will be centered.

Imagine that you have two images that need to be the same size. They are the same width but one is taller than the other. The taller one has a bunch of sky in the top portion and it can be made the same height as the other by cropping off the top. First check the images with identify to see exactly what needs to be done.

The tall one is 84 pixels taller than the ok one. It needs the top 84 pixel rows cropped off. This is done with the following.

Note that this converted to jpg at the same time. ImageMagick tends to need a source and target name even if they are both the same file.

Here is an example of where I wanted to cut off the bottom 7 pixels off of an image.

This will change an image with transparent regions into one with a color (white here) background.

If you have a bunch of such images that you need to fix in place, this works.

Since the border feature actually creates a separate image of the border and then overlays it, you can use this to provide a background for transparent PNGs and GIFs and other formats that possibly have transparent backgrounds. I find this most useful when exporting to a bitmap from Inkscape. In Inkscape, everything looks fine, but when you go to a browser or viewer, it turns out that the lack of background is quite annoying. You could put a big white rectangle at the lowest level of your Inkscape vector model before exporting, but this fix from ImageMagick may be easier since all of your images can be done in one quick operation.

Imagemagick is good at doing screen captures. I prefer it to fancy GUI things because it can be properly scripted and controlled.

Find the window ID with xwininfo.

Also for the whole screen use something like this.

Note that this just captures the first monitor.

Often I want a frame of a video I’m watching. I have the video paused in the player and I just want its entire window. I use this alias to do the job.

Line breaks are obviously not really ideal there. And don’t forget to use .png if you need some kind of image that is not photographic.

Another common thing I want is to grab some part of the screen — maybe it’s a section of video I’m watching or a Blender sub window, but just using some kind of geometry will not work. The thing to do here is to get a terminal and size it over the region of interest. Then use xwininfo and then move the terminal out of the way. Then you can use a command like this (repeatedly if that’s helpful).

Remember how YouTube used to have those little text overlays you could add? Then people started using them only for obnoxious advertising. Well, I used them so that my videos didn’t need a stupid obnoxious sound track (I could explain in text the important things). I’ve been looking for a way to get that functionality back and I think that ImageMagick can do it.

Check the annotate features.

Here is an example of putting a caption on a video frame before assembling it.

Maybe you need to annotate an image or some video frames with a reference grid or marker. See the next section for some tips on that.

ImageMagick can do much fancier things. For example, if you need some lines superimposed over an image. In this example, I wanted some lines superimposed over some dashcam lane lines (red on the port side, green on starboard). I also wanted to crop so just the road was visible.

ImageMagick does a good job of quickly putting borders around images. Here’s an example that puts a red border around an image where the border on the left and right is 10 pixels and the border on the top an bottom is 5 pixes.

If you’re trying to mix parts or elements of two (sometimes more) images, you can use ImageMagick’s composite command. (There are also often ways to do the same thing with convert -composite but worrying about that just adds complexity usually.)

Let’s say you’re labeling dashcam images for segmentation training and you have a hand drawn png of the road or whatever feature you’re looking for — how can you show this mask lightly superimposed over the original image?

Change that dissolve value to vary how transparent the mask is.

What if in the same situation, your dashcam footage always had the hood of the car? Maybe when you were annotating it by hand it was convenient to just bucket fill "road" all the way to the bottom of the frame but now you want to go back and remove all the parts that are static in every image. Here’s how that can be done.

The way this works is the composite command layers the three images sequentially as loaded.

Here’s a more complex example. The situation is there is a 2 color mask image of blue and white, mask.png. This could have some features marked up as detected by some neural net or something. Then there is the original photo, scene.jpg. This will highlight the original photo with red where the scene formerly was white and simply show the original photo scene where it was blue.

It does these steps.

If you work with sets of images it can be nice to have a way to visualize the whole collection at a glance. A "montage" is an image composed of a grid of other images. Here are settings I use the most.

Note that the geometry will control the geometry of the inset images. The +1+1 component of the geometry specification controls how much padding there is around the images. Other options of interest.

Full official details on montage.

Here’s an example.

Here’s how I made a video of two sets of images shown side by side.

The original image sizes here were 350x350 and the montage is 700x350.

Maybe you have a zillion big images and you want a montage to roughly show what this archive contains. I have had trouble with running out of memory. There apparently are ways to limit memory usage. Some of that is described here. As the comments suggest, the CLI options didn’t work for me. I didn’t try the config file. I was able to process 1000 photos by converting them to their thumbnail size into another directory as an intermediate step.

Here’s a way to summarize the process of making the montage in two steps.

Even doing it this way can cause memory problems if your images are large enough. Remember, that ImageMagick is keeping the full image (not compressed) in memory when it works on it.