Azure document translation limitation on the docs context in garment industry with sketch loge drawing image

Question

We are using Azure Document Translation for Chinese→Khmer garment tech-pack PDFs. For digital PDFs, Khmer font sizes are inconsistent due to font substitution, and short text runs are sometimes detected as English unless punctuation is added. For scanned PDFs, OCR-based rendering leaves residual source glyph pixels (“black dots”) because white background masks are tightly bounded to OCR boxes. Logos and company names lose typography even when glossary entries map source=text → target=text. We understand these are OCR/rendering limitations. Please confirm:

1. There is no current control for mask padding / bounding-box expansion.
2. There is no region-level exclusion for logos/figures in Document Translation..
3. Glossary replacement always re-renders text and cannot preserve original fonts.

Answer 1

Hello YM Set Sophy,

Thank you for reaching out to Microsoft Q&A,

You’re not missing anything here, what you’re seeing is exactly how the service behaves today, especially with documents like garment tech packs that mix text, layouts, logos, and drawings.

Let me walk through this in a straightforward way.

Why all of this is happening

Document Translation isn’t editing your PDF in place. It’s more like:

• extract text (either from PDF text layer or OCR)
• translate it
• then redraw the document

That last step re-rendering is where most of these issues come from. The service tries to recreate the layout, but it doesn’t preserve everything exactly the way it was originally designed.

That’s why layout-heavy documents (like tech packs) tend to show more problems compared to plain text PDFs.

On your specific points

You’re correct on all three, and here’s how they play out in practice.

Mask padding / bounding boxes (the “black dots” issue) What you’re seeing in scanned PDFs is a side effect of OCR. The service detects text regions and places a white mask over them before writing translated text. Those masks are very tightly fitted to the detected text.

There’s currently no way to tell it:

• “make the mask slightly bigger”
• or “clean the background more aggressively”

So any tiny part of the original glyph that falls outside that box shows up as those residual dots.

No way to exclude logos or drawings Right now, everything that looks like text is treated the same way.

There’s no option to say:

• “ignore this area”
• “don’t translate this logo”
• “skip this diagram”

So if your logo or branding contains text, the service will:

• detect it
• translate it
• redraw it

Which is why you lose the original styling.

Glossary does not preserve fonts or styling Glossary only controls what text becomes, not how it looks.

Even if you map something like brand name - same brand name

the service still removes the original text and redraws it using available fonts. It doesn’t reuse the embedded font from the PDF.

So typography loss is expected.

The other things you mentioned

Khmer font inconsistency This is usually font substitution during rendering. If the exact font isn’t available or supported in the output pipeline, it falls back to something close which can change size and spacing.

Short text being detected as English Also expected. Very short strings don’t give the language detector enough context. Adding punctuation or surrounding context helps, which is why your workaround improves it.

So what does this mean in reality?

There are a few hard limitations today:

• You can’t control OCR masking behavior
• You can’t exclude specific regions (like logos or sketches)
• You can’t preserve original fonts or styling through glossary

So this isn’t something you can fix by tweaking settings it’s just how the service is designed right now.

What people usually do for documents like yours

For garment tech packs, most teams don’t rely on a single “translate the whole PDF” step. They break it into a small pipeline.

Before translation

• Flatten or outline logos (so they become shapes, not text)
• Remove or isolate drawings/branding if possible

During translation

• Use Document Translation for the actual content (descriptions, specs, tables)

After translation

• Put logos back
• Reapply fonts if needed
• Or overlay translated text onto a clean version of the original

For scanned documents specifically, some teams also run their own OCR first (with better control over masking), clean the image, and then pass it to translation.Please refer this

Azure Translator document translation overview https://learn.microsoft.com/azure/ai-services/translator/document-translation/overview

Azure Translator in Foundry Tools Transparency Note (limitations section) https://learn.microsoft.com/azure/foundry/responsible-ai/translator/transparency-note

Known issues for Azure AI Translator https://learn.microsoft.com/en-us/azure/ai-services/translator/reference/known-issues

I Hope this helps. Do let me know if you have any further queries.

---

If this answers your query, please do click Accept Answer and Yes for was this answer helpful.

Thank you!

Answer 2

1. There is no documented control in Azure Document Translation to adjust OCR mask padding or expand bounding boxes around recognized text. OCR behavior (including how tightly background masks follow text boxes and how residual pixels appear) is an inherent limitation of the underlying OCR and rendering pipeline and is not configurable from the Document Translation API.
2. There is no documented region-level exclusion mechanism in Azure Document Translation to skip logos, figures, or specific layout regions. Document Translation operates over the supported document formats as a whole and, for PDFs and images, uses OCR to extract text; it does not expose controls to define protected or non-translated regions at layout level.
3. Glossaries in Document Translation control term-level translation behavior (for example, enforcing specific translations, or enforcing “no translation” by using the same source and target text), but they do not preserve original typography or fonts. As documented, glossaries are for “context-specific terminology,” “no translation,” and “specify translations for ambiguous words,” not for font or layout preservation. Font style retention is handled by the document processing pipeline, which may substitute fonts where needed; glossary replacement does not provide a way to keep original fonts or logo typography.

---

References:

• Document Translation frequently asked questions
• Capabilities and limitations of optical character recognition
• What is Azure Translator document translation in Foundry Tools?
• Use glossaries with Document translation
• Overview of document translation