Filtering Word-Generated XObject Artifacts in DOCX PDFs

[MSY99]

❗ Filtering Word-Generated XObject Artifacts in DOCX PDFs

Hi,
When extracting images from PDFs using PyMuPDF, I am seeing a large number of unwanted XObjects in PDFs generated from DOCX files—especially when the document contains shapes, charts, or WordArt.

These DOCX PDFs often include:

• monochrome or page-sized background bitmaps
• repeated mask objects (has-mask = true)
• fallback rasterizations of vector elements

Even if the page visually contains only one actual image, page.get_images(full=True) may return many XObjects.

---

📌 What I’m trying to understand

1. Which XObject properties are most reliable for identifying Word-generated artifacts?

Examples I am examining:

• repeated digest
• very large or very small bounding boxes
• color space (DeviceGray vs. DeviceRGB)
• Filters (DCTDecode / Flate / JPX)
• mask usage (has-mask = true)
• transform matrix patterns

Is there a commonly recommended approach in PyMuPDF to distinguish:

real content images vs. layout/background fallback images
generated by Word?

---

📌 Minimal test code

def filter_images_with_debug(page, doc, min_width=30, min_height=30, exclude_grayscale=True):
    images = page.get_images(full=True)
    filtered = []

    for img in images:
        xref = img[0]

        # Show image_info
        try:
            info = page.get_image_info(xref)
            if info:
                for meta in info:
                    print("bbox:", meta.get("bbox"))
        except Exception as e:
            print("image_info error:", e)

        # Extract raster
        try:
            base = doc.extract_image(xref)
            if not base or "image" not in base:
                continue

            pil = Image.open(BytesIO(base["image"]))

            if pil.width < min_width or pil.height < min_height:
                continue
            if exclude_grayscale and is_grayscale_image(pil):
                continue

            filtered.append({
                "xref": xref,
                "width": pil.width,
                "height": pil.height,
                "format": base.get("ext", "png")
            })

        except Exception as e:
            print("extract error:", e)

    return filtered

---

📌 Question for the PyMuPDF team

Are there specific XObject dictionary keys or structural PDF properties that reliably distinguish Word-generated artifacts from meaningful images?

Or alternatively:

Is operator-level inspection (Do, BI, ID) the recommended strategy for this case?

Any guidance would be greatly appreciated.

[JorjMcKie]

Sorry I'm afraid we are no wiser than you here.
What I've observed though is that there are significant differences in internal PDF structures depending on the export method Word -> PDF: Using LibreOffice will create a very different internal PDF structure compared to Word itself, which in turn is very different from the results of every Print-To-PDF software.

If you have access to the original Office files, try to stick with to one software doing the import before you invest too much effort here.

[MSY99]

Thanks for the helpful responses.
After running additional tests on my side, I also observed that the PDF conversion engine used to generate the file (e.g., PDFium, MS Office, LibreOffice, etc.) has a significant impact not only on image extraction, but on overall PDF element extraction behavior (images, tables, text, layout).

Given this, the most practical approach seems to be:

• extracting the PDF producer / generator information from the PDF metadata, and
• branching the image / table / text extraction logic based on the detected conversion engine.

It turned out to be a trickier problem than I initially expected.

Anyway, thank you very much for your help!