Inline image EI markers only divided by a few renderable characters fail to detect the correct one

[AbdiHaryadi]

I found a document which raised a PdfReadError. This error was unexpected because the document can be read properly with, for example, Adobe Acrobat Reader DC.

Environment

I used Google Colab.

$ python -m platform
Linux-6.1.123+-x86_64-with-glibc2.35

$ python -c "import pypdf;print(pypdf._debug_versions)"
pypdf==6.0.0, crypt_provider=('cryptography', '43.0.3'), PIL=11.3.0

Code + PDF

This is a minimal, complete example that shows the issue:

from pypdf import PdfReader
reader = PdfReader("/content/UNIQ_LAP_TAHUNAN_2021-69.pdf")

for page in reader.pages:
    text = page.extract_text()
    print(text)

Access the PDF here.

Traceback

This is the complete traceback I see:

---------------------------------------------------------------------------

PdfReadError                              Traceback (most recent call last)

[/tmp/ipython-input-3995888517.py](https://localhost:8080/#) in <cell line: 0>()
      3 
      4 for page in reader.pages:
----> 5     text = page.extract_text()
      6     print(text)

[/usr/local/lib/python3.12/dist-packages/pypdf/_page.py](https://localhost:8080/#) in extract_text(self, orientations, space_width, visitor_operand_before, visitor_operand_after, visitor_text, extraction_mode, *args, **kwargs)
   2036             orientations = (orientations,)
   2037 
-> 2038         return self._extract_text(
   2039             self,
   2040             self.pdf,

[/usr/local/lib/python3.12/dist-packages/pypdf/_page.py](https://localhost:8080/#) in _extract_text(self, obj, pdf, orientations, space_width, content_key, visitor_operand_before, visitor_operand_after, visitor_text)
   1719         extractor.initialize_extraction(orientations, visitor_text, cmaps)
   1720 
-> 1721         for operands, operator in content.operations:
   1722             if visitor_operand_before is not None:
   1723                 visitor_operand_before(operator, operands, extractor.cm_matrix, extractor.tm_matrix)

[/usr/local/lib/python3.12/dist-packages/pypdf/generic/_data_structures.py](https://localhost:8080/#) in operations(self)
   1404     def operations(self) -> list[tuple[Any, bytes]]:
   1405         if not self._operations and self._data:
-> 1406             self._parse_content_stream(BytesIO(self._data))
   1407             self._data = b""
   1408         return self._operations

[/usr/local/lib/python3.12/dist-packages/pypdf/generic/_data_structures.py](https://localhost:8080/#) in _parse_content_stream(self, stream)
   1297                     peek = stream.read(1)
   1298             else:
-> 1299                 operands.append(read_object(stream, None, self.forced_encoding))
   1300 
   1301     def _read_inline_image(self, stream: StreamType) -> dict[str, Any]:

[/usr/local/lib/python3.12/dist-packages/pypdf/generic/_data_structures.py](https://localhost:8080/#) in read_object(stream, pdf, forced_encoding)
   1474     stream.seek(pos)
   1475     read_until_whitespace(stream)
-> 1476     raise PdfReadError(
   1477         f"Invalid Elementary Object starting with {tok!r} @{pos}: {stream_extract!r}"
   1478     )

PdfReadError: Invalid Elementary Object starting with b'\xff' @1157515: b'\xa0\x9d0\nh\x08 \xa9z0\x00\x01~~ \t\x00EI\n\xff\nEI Q\nQ\nq\n4205.52 6751.53 2.47266 576.254 re W n\nq 0 576.25'

[stefan6419846]

Thanks for the report. This most likely is a limitation of PDF < 2.0 and how to correctly parse inline images. We have some heuristics to check whether their end marker EI is followed by binary data or not. In this specific case, it apparently fails as \n\xff\n does not look like binary content.

Adding

    if operator == b"EI":
        return True

to the corresponding function seems to solve this issue, although there still are limitations for text retrieval due to

/c65/c110/c103/c103/c111/c116/c97/c32/c75/c111/c109/c105/c116/c101/c32/c65/c117/c100/c105/c116

most likely not being really useful for you.

I still need to investigate whether adapting the inline image detection logic like indicated has any side effects before pypdf getting an official fix.

[stefan6419846]

I had another thought about this, and the proposed approach will not work as we could have even more EI sequences where only one is valid. I am unsure whether we can really solve this in a reliable fashion without too much overhead, as we would (in theory) have to check all possible ends for parsing errors and go with the best one of them.