How to rebuild a page instance and put it back to the pdf with the results of page.get_text("json")?

[lidh15]

the case is simply from I need to remove some blocks in some pages of the pdf file. It's not difficult to get the json string and manipulate the blocks in the dict, but after that I want to recreate the pdf file, is there any API available?

[JorjMcKie]

Don't take the JSON string, take the "dict" format instead. JSON is derived from it.

You cannot write back to the same page - instead make a new page in a new PDF and write (a subset of) the contents of the dictioanry to it.
There is an example set of scripts for font replacement, where you can probably learn a lot from.

[lidh15]

So another question is that how to find which block is on the top? I was able to remove a block, but I need to know which block is on the top thus gonna be shown and then I can remove the others.

[lidh15]

And I have another curious topic, how to transfer the image block we get into a PIL image instance like return by PIL.Image.open?

[JorjMcKie]

If you modify the JSON output of font replacement like this, the resulting PDF will look good ... and exhibit text that was hidden under those inline images 😎.

[
  {
    "oldfont": [
      "\u008b\u00cc\u00e5,Bold",
      "\ufffd\ufffd\ufffd,Bold"
    ],
    "newfont": "cjk",  # the built-in universal font supporting Chinese
    "info": "Not embedded!"
  },
  {
    "oldfont": [
      "\u00bf\u00ac\u00cc\u00e5_GB2312,Bold",
      "\ufffd\ufffd\ufffd\ufffd_GB2312,Bold"
    ],
    "newfont": "cjk",
    "info": "Not embedded!"
  },
  {
    "oldfont": [
      "\ufffd\ufffd\ufffd\ufffd\ufffd\ufffd\ufffd,Bold",
      "\u00bb\u00aa\u0087\u00bf\u00ac\u00cc\u00e5,Bold"
    ],
    "newfont": "cjk",
    "info": "Not embedded!"
  }
]

[lidh15]

Thank you, I followed this and it worked, but I can only open the modified pdf with Chrome, when I tried to open it with adobe pdf reader it said that there were errors in the page and adobe acrobat cannot render the page correctly. The reader somehow opened it but the text written by "fitz.text_writer" and the images inserted by "page.insert_image" were not seen. Is there any clues for the reason why this happened?

[JorjMcKie]

Try to do it again with the non-hacky method (using redactions). It is "official" and not that error-prone.
My Adobe has no complaints, so you probably made some kind of error.
Otherwise I know that Adobe is very picky.

[lidh15]

Is that related to the arguments of doc.save()?
I keep using the settings from the repl-font.py, and I'm wondering how it makes differences when I have different value for "deflate" and "garbage"? What arguments combination can provide the smallest file size?

[JorjMcKie]

What arguments combination can provide the smallest file size?

There is a save alias called "easy save", ez_save() with some different default settings garbage=3, deflate=True. Which is a good balance between size reduction and save speed. Use garbage=4 to achieve the top compression (with MuPDF).

[JorjMcKie]

how to find which block is on the top?

When MuPDF builds its TextPage object (which I wrap with the same-named Python object), it applies some heuristics when it creates its block-line-span-character hierarchy.
This process leaves the original sequence of the underlying PDF painting instructions intact - at least roughly.

Some minor deviations may happen depending on the flags parameter, which may cause space characters to be generated or general whitespace be replaced by one or more spaces.

So, in the generated dictionary get_text("dict", ...), the sequence of the text pieces should reflect the sequence of the corresponding original painting commands. If text bboxes overlap, one might conclude that a later one overlaps an earlier one.

I also have page.get_bboxlog() which is a list of rectangles flagged with the rectangle "type" (text, drawing, image). Their sequence indeed is the painting sequence.
This is a more complete picture - i.e. not only regarding text, but all sort of objects potentially overlapping each other.