Rotating PDF based on orientation of JSON

[orlandito24]

I was wondering if using the orientation provided by the JSON, there was a way I could orientate the PDF I'm reading into DocumenFile.from_pdf().

I'm trying to capture the coordinates of lines/words, but on rotated PDF's there are too many coordinates in play so I want the PDF to be straightened out so every line/word has two coordinates. Let me know if there is a better way to go about this. thank you!

[felixT2K]

Hi @orlandito24 👋🏼,

We are on it currently: #1336
Additional with #1364 this approach will be much more robust for rotations between -90 and 90 degrees.
So it will be available with the next release.

After the release you can pass assume_straight_pages=False detect_orientation=True straighten_pages=True to the ocr_predictor instance, internally we will straighten the pages based on the page orientations in this case it will return "straight" boxes.

For the moment you can use export_as_straight_boxes=True if assume_straight_pages is set to False (NOTE: this works atm only with PT: db_resnet50_rotation TF: linknet_resnet18_rotation as detection model)

For visibility: @odulcy-mindee

[felixT2K]

Nice 👍🏼

[orlandito24]

@felixT2K There is one thing I'm realizing, not sure if it is possible but the render seems to be sorting blocks and printing the text inside. I would want more of a sort where it would print the lines sorted from top left to right down. Right now the floating PDF DOC PNG are being printed after the Digital Documents, Even thought the former is closer to the top left.

[felixT2K]

@orlandito24 I think that's tricky. Are the single words (PDF, DOC, PNG) sorted into one block or each of these words into one block element ?

[felixT2K]

If it's the first case you could do your own sorting on the coordinates of the block elements

[orlandito24]

Hi @felixT2K, I was wondering what would the best ocr_predictor combo for files that are straight and I would want straight boxes around the text?