OCR failure on certain PDF pages

[vikasr111]

Bug description

I am trying to perform OCR using DocTR on a PDF document. I have noticed that the OCR for the whole document is failing because of one page in the document. When I ran that individual one-page pdf, I got the full error log. Here's the PDF:
po-r.pdf

Here's the error log:

Traceback (most recent call last):
  File "/app/main.py", line 25, in <module>
    main(args.filename)
  File "/app/main.py", line 12, in main
    doctr_onnx(filename)
  File "/app/doctr_onnx.py", line 394, in doctr_onnx
    result = predictor(docs)
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/predictor/predictor.py", line 132, in __call__
    word_preds = self.reco_predictor([crop for page_crops in crops for crop in page_crops], **kwargs)
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/recognition/predictor/base.py", line 67, in __call__
    processed_batches = self.pre_processor(crops)  # type: ignore[arg-type]
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/preprocessor/base.py", line 105, in __call__
    samples = list(multithread_exec(self.sample_transforms, x))
  File "/usr/local/lib/python3.10/site-packages/onnxtr/utils/multithreading.py", line 47, in multithread_exec
    results = map(lambda x: x, tp.map(func, seq))  # noqa: C417
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 367, in map
    return self._map_async(func, iterable, mapstar, chunksize).get()
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 774, in get
    raise self._value
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 125, in worker
    result = (True, func(*args, **kwds))
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 48, in mapstar
    return list(map(*args))
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/preprocessor/base.py", line 70, in sample_transforms
    x = self.resize(x)
  File "/usr/local/lib/python3.10/site-packages/onnxtr/transforms/base.py", line 58, in __call__
    img_resized = Image.fromarray(img).resize(tmp_size, resample=self.interpolation)
  File "/usr/local/lib/python3.10/site-packages/PIL/Image.py", line 2200, in resize
    return self._new(self.im.resize(size, resample, box))
ValueError: height and width must be > 0

I printed the docs of DocTR to debug further and here's the result:

Number of pages in po-r.pdf: 1
Pages: [array([[[255, 255, 255],
        [255, 255, 255],
        [255, 255, 255],
        ...,
        [255, 255, 255],
        [255, 255, 255],
        [255, 255, 255]],

       [[255, 255, 255],
        [255, 255, 255],
        [255, 255, 255],
        ...,
        [255, 255, 255],
        [255, 255, 255],
        [255, 255, 255]],

       [[255, 255, 255],
        [255, 255, 255],
        [255, 255, 255],
        ...,
        [255, 255, 255],
        [255, 255, 255],
        [255, 255, 255]],

       ...,

       [[255, 255, 255],
        [255, 255, 255],
        [255, 255, 255],
        ...,
        [255, 255, 255],
        [255, 255, 255],
        [255, 255, 255]],

       [[255, 255, 255],
        [255, 255, 255],
        [255, 255, 255],
        ...,
        [255, 255, 255],
        [255, 255, 255],
        [255, 255, 255]],

       [[255, 255, 255],
        [255, 255, 255],
        [255, 255, 255],
        ...,
        [255, 255, 255],
        [255, 255, 255],
        [255, 255, 255]]], dtype=uint8)]

Code snippet to reproduce the bug

from doctr.models import ocr_predictor
from doctr.io import DocumentFile

# load image
pdf_path = "./data/invoice/po-r.pdf"

# extracting text from input image using docTR
docs = DocumentFile.from_pdf(pdf_path)

# load model
predictor = ocr_predictor(
    det_arch="db_mobilenet_v3_large",
    reco_arch="parseq",
    pretrained=True,
    preserve_aspect_ratio=False,
    symmetric_pad=False,
    )

predictor.det_predictor.model.postprocessor.bin_thresh = 0.35
predictor.det_predictor.model.postprocessor.box_thresh = 0.3


result = predictor(docs)

# display ocr boxes
result.show(docs)

Error traceback

Here's the error log:

Traceback (most recent call last):
  File "/app/main.py", line 25, in <module>
    main(args.filename)
  File "/app/main.py", line 12, in main
    doctr_onnx(filename)
  File "/app/doctr_onnx.py", line 394, in doctr_onnx
    result = predictor(docs)
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/predictor/predictor.py", line 132, in __call__
    word_preds = self.reco_predictor([crop for page_crops in crops for crop in page_crops], **kwargs)
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/recognition/predictor/base.py", line 67, in __call__
    processed_batches = self.pre_processor(crops)  # type: ignore[arg-type]
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/preprocessor/base.py", line 105, in __call__
    samples = list(multithread_exec(self.sample_transforms, x))
  File "/usr/local/lib/python3.10/site-packages/onnxtr/utils/multithreading.py", line 47, in multithread_exec
    results = map(lambda x: x, tp.map(func, seq))  # noqa: C417
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 367, in map
    return self._map_async(func, iterable, mapstar, chunksize).get()
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 774, in get
    raise self._value
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 125, in worker
    result = (True, func(*args, **kwds))
  File "/usr/local/lib/python3.10/multiprocessing/pool.py", line 48, in mapstar
    return list(map(*args))
  File "/usr/local/lib/python3.10/site-packages/onnxtr/models/preprocessor/base.py", line 70, in sample_transforms
    x = self.resize(x)
  File "/usr/local/lib/python3.10/site-packages/onnxtr/transforms/base.py", line 58, in __call__
    img_resized = Image.fromarray(img).resize(tmp_size, resample=self.interpolation)
  File "/usr/local/lib/python3.10/site-packages/PIL/Image.py", line 2200, in resize
    return self._new(self.im.resize(size, resample, box))
ValueError: height and width must be > 0

Environment

Python - 3.11.7

Deep Learning backend

Torch

[felixdittrich92]

Thanks for reporting 👍

Would it be possible to share such an pdf that we can reproduce the issue ?
And i see it's OnnxTR related ^^

/usr/local/lib/python3.10/site-packages/onnxtr/transforms/base.py

[vikasr111]

Here's the pdf link: https://github.com/user-attachments/files/17991370/po-r.pdf

I ran it through both DocTR and OnnxTR and ran into same problem.

[felixdittrich92]

Hey @vikasr111 👋,

I tested it with both docTR and OnnxTR without success to reproduce the bug ..Could you try to provide the absolute path to the pdf ?

I used the same args as provided in your snippet - only changed the pdf path to absolute

[felixdittrich92]

result.show(docs) is outdated btw it's only result.show() since v0.9 if i remember correctly

docTR:

OnnxTR:

[vikasr111]

@felixdittrich92 That's odd. When I run this directly in my system it works fine. But when I run the same code in docker container I get this error.

Will investigate more as this error is coming from Pillow where it's getting blank image for the page.

[felixdittrich92]

Maybe an issue with pypdfium2 ?
What do you use as base image in your dockerfile ?

[vikasr111]

@felixdittrich92 I use python:3.11-slim-bullseye

Here's my full Dockerfile:

# Base Stage: Install common dependencies
FROM python:3.11-slim-bullseye

# Set the working directory in the container to /app
WORKDIR /app

# Set PYTHONPATH to include the /app directory
ENV PYTHONPATH="${PYTHONPATH}:/app"

# Add the rest of the current directory contents into the container at /app
COPY . .

# Install dependencies for OCR
RUN apt-get update \
    && apt-get install --no-install-recommends -y libgl1-mesa-glx libglib2.0-0  build-essential \
    # Install your Python packages that require compilation here
    && pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu \
    && pip install "python-doctr[torch]" \
    && pip install --no-cache-dir "onnxtr[cpu]" \
    && pip install numpy scikit-learn matplotlib pymupdf \
    # Cleanup: Remove build-essential and cleanup apt cache in the same command
    && apt-get purge -y build-essential \
    && apt-get autoremove -y \
    && rm -rf /var/lib/apt/lists/* /root/.cache/pip

I found some lead. When I set symmetric_pad=False it works fine but when I set it to True it starts giving error for this particular PDF only.

[felixdittrich92]

@vikasr111 quickly tested: 😅
Does also not work on your local machine ?

[felixdittrich92]

Also created a fresh environment that all sub-packages are up-to-date