Using Imagemagick to convert PDFs having rasters in them results in white backgrounds behind the sections covered by the rasters - imagemagick

Question

I tried to use ImageMagick (v6.8.9-9 Q16) to convert a PDF containing a PNG file embedded in it to a PNG file.
The original PNG file had a transparent background. In the PDF too it appears fine. But in the PNG obtained after conversion, the area originally occupied by the PNG in the PDF has a white background. Please see the links for more clarity.
The command I ran is as follows:
convert -colorspace sRGB dice.pdf converted_dice.png
I also tried setting the -transparent white switch but it ends up taking out whites that were actually required in the final image.
Are there any extra switches or parameters to pass to convert in order to get rid of just this white background?

Answer 1

Kurt already explains the whole thing in great detail. So here is just how to assemble an image with ImageMagick after running it through pdfimages -png
pdfimages -png my.pdf my
This resluts in two files
identify my-0*png
my-000.png PNG 360x310 360x310+0+0 8-bit sRGB 256c 3.3KB 0.000u 0:00.000
my-001.png PNG 360x310 360x310+0+0 8-bit sRGB 256c 9.44KB 0.000u 0:00.000
my-001.png is the image labeled smask in pdfimages -list. To reassemble the image back to it's original form use -compose CopyOpacity with the ImageMagick command composite
composite -compose CopyOpacity my-001.png my-000.png my-reassembled.png
See also http://www.imagemagick.org/Usage/masking/#masks for more information.

Answer 2

Your approach to this task cannot work.
The command you used will convert the complete letter-sized PDF page (612 x 792 pt) into a PNG image.
However, the original size of the image embedded in the PDF page (612 x 792 pt) is 800 x 600 pixels. This can be seen by running pdfimages -list:
pdfimages -list dice.pdf
page num type width height color comp bpc enc interp object ID x-ppi y-ppi size ratio
----------------------------------------------------------------------------------------
1 0 image 800 600 rgb 3 8 image no 12 0 72 72 277K 20%
1 1 smask 800 600 gray 1 8 image no 12 0 72 72 50.1K 11%
So this is the first problem when converting the PDF page: it does not give your the correct size of the contained images.
The second, more fundamental problem however is: any image you get from converting a PDF page is the combination of all PDF objects overlayed on each other as they are from the page area. (Of course you could crop only part of the page -- but this gives you likewise the combination of all PDF objects from the cropped area...). The results of this you've encountered when you tried to convert all white pixels into transparent ones: since the originally different objects are merged into one representation of pixels, you can no longer discriminate between them as required.
You should take a different approach and use a different tool to extract the image: use pdfimages (the tool used above with the -list parameter to display image properties from the PDF's pages). As you can see, there are two images list: one is an RGB raster image, the other is a grayscale raster image, dubbed as type smask (softmask).
Here is a command to extract both images as PNG:
pdfimages -png dice.pdf dice-images
This will extract the two:
dice-images-0000.png (a color image)
dice-images-0001.png (a grayscale image)
(Note: Only very recent versions of pdfimages, the Poppler version, will let your extract the images as PNG. Within the PDF there is no such thing as PNG. There are only raster data, compressed with different methods. Older versions will only be able to extract images as PPM or PNM. This does not have any influence on what I describe below. Even if you extract PPM/PNM images, these two files can still be processed as described below...)
Below is a side-by-side, scaled-down montage of the two:
As you can see, the image itself does not have a transparent background, but a white one. (It does not have an Alpha channel.) Within the PDF format, these two images are used in combination to create transparent areas:
what appears completely black in the softmask (right) means: this pixel of the real image (left) is meant to be fully transparent.
what appears completely white in the softmask (right) means: this pixel of the real image (left) is meant to be fully opaque.
what appears in a shade of gray in the softmask (right) means: this pixel of the real image (left) is meant to be partially transparent (in line with its level of gray/black).
To combine these two files (color image and grayscale softmask) back into one PNG with transparency, you can employ ImageMagick now...