Document.extract_image seems to extract image png mirrored in some cases

[p1-dta]

Hi,

I have a pdf with only a png in it, but when I get try to extract this png, it's mirrored.

Document.extract_image return a dictionary with, for the value "image", the image but mirrored.
This happen only for one file, works fine for many more.

I have no clues, neither do my colleagues.
I don't know how to prevent it or how to detect a mirror to mirror it again.

[JorjMcKie]

Guilty is the PDF creator, who decided to do this!
You have the option to extract the "transformation matrix" that was used to make an image fit into its bbox on the page (scaling, mirroring, rotating).
Analyzing this matrix will give you the desired information.
Read the documentation here for details.
This code snippet may help with such a matrix analysis:

import fitz

msg = ""
m = fitz.Matrix()
if m.b * m.c == 0:  # means rotation by 0, 180, or flippings
    if m.a * m.d > 0:  # same sign -> no flippings
        if m.a < 0:  # so both, a and d are negative!
            msg = "rot 180"
        else:
            msg = "nothing"
    else:  # we have a flip
        if m.a < 0:  # horizontal flip
            msg = "left-right"
        else:
            msg = "up-down"
elif abs(m.b) == abs(m.c):
    if m.b * m.c < 0:
        if m.b > 0:
            msg = "rot 90"
        else:
            msg = "rot 270"
    else:
        if m.b > 0:
            msg = "up-down, rot 90"
        else:
            msg = "rot 90, updown"
else:
    error = True

if error:
    print("unsupported matrix")
else:
    print(msg)

[p1-dta]

Thanks, I mark the question as solved if this solve my issue.

[p1-dta]

I get the matrix with _, mat = page.get_image_bbox(page.get_images(full=True)[0], transform=True) but mat == Matrix(850.0, -0.0, -0.0, 1169.0, -0.0, 1169.0) and the matrix analyzer return "nothing".

I compared 2 images, one flipped up-down (should be d = -1), one normal, and here is the respective matrix :

flipped up-down: Matrix(850.0, -0.0, -0.0, 1169.0, -0.0, 1169.0)
normal: Matrix(402.5350036621094, -0.0, -0.0, 841.9999389648438, 96.23249816894531, -0.0)

I can't tell the difference without looking the result with my eyes.

I checked:

• https://pymupdf.readthedocs.io/en/latest/page.html#Page.get_image_bbox
• https://pymupdf.readthedocs.io/en/latest/document.html#Document.get_page_images
• https://pymupdf.readthedocs.io/en/latest/matrix.html#examples
  and [a, b, c, d, e, f] attributes.

[JorjMcKie]

What do the EXIF data of the images say?

[JorjMcKie]

I need a file example.

[p1-dta]

Unfortunately, I can't share this document. Is there anything else I can do ?

[JorjMcKie]

No, I did understand your issue, but the only way for further investigation is looking at the file itself.

[p1-dta]

I'll ask if it's possible, maybe in a few days. Thanks for your help.

[p1-dta]

@JorjMcKie I have an update. I can share the file with you. Do you want me to send it to you on your mailbox ?

[JorjMcKie]

Yes please do so. Yesterday I found an error in computing the transformation matrix via `get_image_bbox()`. This may affect you. You can try `get_image_rects()` instead. This is always correct. It returns a list of tuples `(rect, matrix)`. One tuple for each occurrence of the same image on the page. You probably will find the correct answer with this alternative. Von meinem iPhone gesendet Am 6/9/21 um 05:08 schrieb Dorian Turba ***@***.***>:  @JorjMcKie<https://github.com/JorjMcKie> I have an update. I can share the file with you. Do you want me to send it to you on your mailbox ?

[JorjMcKie]

I have opened issue #1244 to address this problem correctly.

[p1-dta]

Do you have any idea about when the fix will be released for get_image_bbox ?

[JorjMcKie]

I am working on it. Should be no later than tomorrow.