Can I convert from pix to PIL without saving to disk?

[alejandrofm]

Hi! I am currently doing this to extract images from PDFs and then feeding those to tesseract-OCR, it seems that I'm wasting, time, quality and not optimizing the code when I write to PNG and read the PNG in the next line, is there any way to feed a pixmap to a Image element?

            pix = page.getPixmap(matrix=mat, alpha=False)  # render page to an image
            saving_name = filepath.split('\\')[-1].replace('pdf', '') # extract name from file
            pix.writePNG("page{}{}.png".format(saving_name, page.number))  # store image as a PNG
            imagen = Image.open("page{}{}.png".format(saving_name, page.number))  # open image
            imagen.show() #show image

Thank you!

[JorjMcKie]

Sure there is!
There are even a number of alternatives to avoid it:

1. You can convert the pixmap to a number of image formats as bytes objects. Use data = pix.getImageData("format") for this. You can choose among several "format" alternatives trading size of the bytes object data for creation speed. Then use img = Image.open(io.BytesIO(data), ...) to open it as a PIL image.

2. Use the pixmap directly as input to create a PIL image: img = Image.frombytes(mode, [pix.width, pix.height], pix.samples). The string "mode" in your case should probably be "RBG".

To improve the OCR detection, you can render the PDF page at higher quality by using a matrix in Page.getPixmap(). E.g. matrix = fitz.Matrix(2,2) will increase resolution by a factor of 2 in each (x, y) direction - thus yileding a 4 times larger pixmap.
But if the PDF contains images (which I assume to be) then that may have limited effect for OCR recognition. If on the other hand the page contains normal text, then there would be no need to use OCR ...

[Aspace2create]

line 146, in keyPressEvent
data=pxm.getImageData("PNG")
AttributeError: 'QPixmap' object has no attribute 'getImageData'

[alejandrofm]

Thank you @JorjMcKie its a single image per page and I'm already using matrix 2,2 with great results.
Any option (1 or 2) or parameter recommendation for the conversion to PIL?
Thanks again!

[JorjMcKie]

Option 2 is about one third faster than option 1 with a similar memory footprint ...
I just measured 10 loops, best of 3: 42.3 ms per loop (option 2) versus 10 loops, best of 3: 64.5 ms per loop (option 1) for a large pixmap with a samples size of about 23 MB.

Option 1 is important for feeding image data to some packages, most notably tkinter. In the latter case you would use pix.getImageData("ppm") as input to tk.PhotoImage in a Python version independent manner.

[alejandrofm]

Thankyou A LOT @JorjMcKie it works great, could I have any problem if the source image is grayscale or b&w? will it "scale" to RGB?
Thanks!

[JorjMcKie]

You are welcome!
No, you can freely choose the resulting colorspace when using getPixmap. If using fitz.csGRAY the pixmap will be smaller of course (only one third of RGB), which may be utterly sufficient for doing the OCR stuff.

[alejandrofm]

I'm VERY happy with the result currently using fitz.csGRAY and mode R , with your help I improved:

• Avoid creating a single a file for each page and writing it to disk.
• Avoid reading file from disk in next line of code.
• Lower size of the file i'm working on.

Thank you a lot!

[JorjMcKie]

close issue?

[rozeappletree]

@alejandrofm

I'm VERY happy with the result currently using fitz.csGRAY and mode R , with your help I improved:

• Avoid creating a single a file for each page and writing it to disk.
• Avoid reading file from disk in next line of code.
• Lower size of the file i'm working on.

Thank you a lot!

Can you please give a code snippet for fitz.csGRAY and mode R

[JorjMcKie]

@rakesh4real @alejandrofm -
Just wanted to let you guys know, that the recent PyMuPDF versions have new methods for direct support of PIL / Pillow - so you no longer need fiddling around with pixmap attributes to achieve this:

• Pixmap.pillowSave(...) saves a pixmap to an image file using Pillow. The arguments are the same as for Image.save(...) in pillow. This makes it possible, to output JPEG images for example ... just everything supported by Pillow.
  Note that this method will be renamed to pixmap.pil_save() in version 1.18.14.
• Similarly, Pixmap.pillowData(...) returns a bytes object using Pillow. Otherwise the same arguments supported.
  Again a rename in version 1.18.14: Pixmap.pil_tobytes().

[canklot]

For those who wants numpy array not PIL.Image

pix = PIL.Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
pix = np.array(pix)

[JorjMcKie]

@canklot - there is an even more direct and much faster way for ndarrays: see here.