Get dimensions of composition, minus whitespace - imagick

Question

I have a series of images (10000's) that are all object-on-white-bg composition. Images are let's say 1000x1000 and objects within are any size. I want to get the actual dimensions of the object, which could be something like 400x500 or 350x520. The rest of the image is just whitespace (but there is some dirtyness to the whitespace e.g. 2% off-white)
I just need analysis of the images; I don't need to export or save data.
QUESTION: Using imagemagick, What's the most performant way to get actual object dimensions on a clean background? I just wonder is there a better imagemagick tool to use than trimImage().
Here is one working method, roughly:
(PHP/Laravel project)
$path = 'images/myTestImage.jpg';
$imagick = new \Imagick(realpath($path));
$originalDimensions = $imagick->getImageWidth()."x".$imagick->getImageHeight();
// Using imagemagick trimImage(), with fuzz value to take care of dirty white pixels.
$imagick->trimImage(5000);
$newDimensions = $imagick->getImageWidth()."x".$imagick->getImageHeight();
Example image:

Answer 1

Here are some ideas. I don't know enough about your environment or images to say which, if any, will work best or be appropriate for you.
You may be able to use the JPEG "shrink-on-load" feature to reduce the demand on I/O, memory and CPU cycles while loading your images - especially if the edges of your shapes are fuzzy by a couple of percent, the reduced size will not make much difference to the finding of the shape outlines. I'll demonstrate by creating a test image to trim:
convert -size 2000x2000 xc:black -fill white -draw "rectangle 100,100 200,200" image.jpg
Now get trim box:
time convert image.jpg -format %# image.jpg info:
112x112+96+96112x112+96+96
real 0m0.128s
user 0m0.803s
sys 0m0.097s
Now do same thing again but with "shrink-on-load" to 1/4 the width and height:
time convert -define jpeg:size=500x500 image.jpg -format %# image.jpg info:
28x28+24+2428x28+24+24
real 0m0.040s
user 0m0.054s
sys 0m0.014s
Note that is 3-4x faster, also that the features (including trim box) are correspondingly reduced.
You may be able to use vips, though you would have to add the corresponding tag to get John's (the author) input. If vips can do this, it may be an excellent option. I think you would need to use find_trim() which is documented here.
You may able to "shell out" using system() to get a tool such as GNU Parallel to use multiple CPU cores to process your images in parallel. It's a quick, easy install. You can just try the following in your Terminal without PHP to get an idea how fast it is:
parallel --bar 'magick -format "%f:%#\n" {} info:' ::: *.jpg
If you have too many images for your command line, you can pump the names into it this way:
find . -name \*.jpg -print0 | parallel -0 --bar 'magick -format "%f:%#\n" {} info:'
You could write some Python that takes a load of filenames to check with OpenCV which will be really fast, and then get GNU Parallel to call that Python with as many filenames as your shell can handle (-X option) to amortise the overhead of starting the Python interpreter over as many images as possible. I mean this:
parallel -X script.py ::: *.jpg
Then your Python would be of the form:
for f in sys.argv[1:]:
image = cv2.imread(... probably greyscale)
cv2.findContours ... à la Rotem https://stackoverflow.com/a/60833042/2836621
Or, rather than findContours() you can probably do it faster as suggested by Divakar here.