Get the transformation applied to an image

[julamb]

Hi. Thanks for a great library with loads of documentation and examples!

I'm trying to follow this example to extract images from a pdf: https://github.com/pymupdf/PyMuPDF/wiki/How-to-Extract-Images-from-a-PDF

My issue is, in some PDFs that are scans of a document, the images seems to be rotated (the export is rotated 90° compared to what I see in the .pdf). I've researched the subject a little and it looks like I could apply the inverse rotation if I could access the transformation matrix (ctm) of every image, but could not find a way to get it from either the xref or the pixmap object. Any hints?

Thanks a lot!
Julien

[JorjMcKie]

You must mean Page._getTransformation().
This method delivers the matrix applied to the page so that its geometry appears as presented in PyMuPDF.
If you want, send me examples with few pages where your problem persists.

[julamb]

Thanks for the quick reply! I've taken a look at the transformation matrix for my page but it doesn't seem to include any rotation. Isn't there a transformation matrix associated to the image element too? Not really familiar with how pdf layouts work.

I've sent you a document by email (unfortunately I could not find a reproduction case that).

[JorjMcKie]

I haven't yet received the document.
Unfortunately, there is more than one way how things can get rotated.

My own page method insertImage rotates by a command in the page's /Contents objects for example. To undo this kind would require writing a de-compiler of the used PDF mini-language. Which you probably do not want to do.

Then, the scanned-in image itself may contain instructions of this sort in its EXIF meta-information. There are Python packages to check this. I used piexif and PIL to do this (another context outside PyMuPDF):

# read orientation value if present (img is a PIL Image)
if "exif" in img.info:
    exif_dict = piexif.load(img.info["exif"])
    orientation = exif_dict["0th"].get(piexif.ImageIFD.Orientation, 1)
else:
    orientation = 1

And of course you might use some heuristics: check whether the image aspect ratio width/height > 1 and the page rectangle width/height < 1?

You could print all / some PDF object definitions and include them here:

xref_count = doc._getXrefLength()
for xref in range(1, xref_count):
    print("===> object:", xref)
    print(doc._getXrefString(xref))

This would allow to check if there is a way to detect this via PDF object inspection.

[JorjMcKie]

You could also try a different method for image extraction than in the example script.
If you have a PDF where every page is completely covered by a scanned-in image:

for pno in range(len(doc)):
    img_list = doc.getPageImageList(pno) # list of all images on page pno: should only have 1 item!
    xref = img_list[0][0]  # the PDF object number of the page image
    img = doc.extractImage(xref)  # extract the image without using Pixmaps
    img_file = open("img-%i.%s" % (xref, img["ext"]), "wb")
    img_file.write(img["image"])
    img_file.close()

This is faster, but more importantly saves the image in its original format, and we can hope this is not rotated.

[JorjMcKie]

@julamb - any reaction on your side?