Why does Amazon Nova OCR still produce “box” symbols (□/■■■) instead of real Chinese text in rebuilt PDFs?

Question

Hi everyone,

I'm building a multilingual OCR → text → PDF rebuilding pipeline using Amazon Bedrock (Nova and Claude models). The PDFs contain mixed Chinese + English text, sometimes scanned, sometimes rasterized.

Problem:
Even when OCR appears to run successfully, some parts of the output still contain “box” symbols such as:

□□□□

� (unknown replacement)

■■■■ (solid block placeholders)

These appear in my final rebuilt.pdf even though the rest of the text is extracted correctly. Pages with dense Chinese text or stylized fonts often produce replacement boxes □■■□ instead of proper characters, even though Nova/Claude successfully detect English on the same page.

Here is the actual PDF causing problems:
https://drive.google.com/file/d/16OU0sjhid7sbV_uNUVPQZoa9fiN0MJTc/view?usp=sharing

Here is my current pipeline and what I already tried:
1. Convert each PDF page → PNG
- Using pypdfium2
- 300 DPI, RGB
- Works fine, images look sharp

2. Send image bytes to Amazon Nova Pro
- Model: amazon.nova-pro-v1:0 (also tested Nova Lite, Titan Embeddings, Claude 3.5 Sonnet/Haiku)

My question is: 
1. Are “box” characters a known fallback for unrecognized Unicode in Nova OCR?
Does Nova use boxes when it cannot decode a glyph? Or is this a downstream encoding issue in my combination of utf-8 strings, python str, reportlab PDF generator

2. Does Nova Pro fully support Chinese OCR?
The marketing describes “multilingual OCR,” but I cannot find: supported language list, performance expectations for CJK fonts, any required headers for multilingual content

3. Should I preprocess images before sending to Nova?

E.g.:
DPI increase (300 → 450 or 600?), grayscale?, contrast/thresholding? sharpen? enlarge regions containing Chinese?

Thank you!

re:Post Agent · Answer

The "box" symbols (□/■■■) in your OCR output when processing Chinese text are likely due to several factors related to both Amazon Nova's capabilities and the nature of the content you're processing.

Amazon Nova does have some limitations with multilingual image understanding. The models can struggle with multilingual images, particularly with complex scripts like Chinese characters. This is especially true when:

1. The text is too small in the images - which appears to be the case in your shared PDF
2. The Chinese characters have complex structures or stylized fonts
3. There's low contrast between text and background
4. The characters might be rare or uncommon in the model's training data

These box symbols typically appear when the OCR model cannot confidently recognize certain characters and outputs placeholder symbols instead.

To improve your results:

**Image preprocessing recommendations:**
- Increase the text size by cropping to relevant sections while preserving context
- Convert images to grayscale to reduce complexity
- Enhance contrast to make text more distinct
- Consider increasing resolution beyond 300 DPI (try 450-600 DPI)
- Apply image denoising if the PDF contains artifacts

**Prompt engineering:**
- Be more specific about languages in your prompt (explicitly mention Chinese and English)
- You might try instructing the model to transliterate characters it's unsure about rather than using placeholders

**Alternative approaches:**
- Process the document in smaller chunks rather than full pages
- Consider using Amazon Textract as an alternative, which might handle multilingual OCR differently

For your specific PDF, the text appears quite small, which is likely a major factor in the OCR challenges. Increasing the relative size of the text regions before processing could significantly improve results.  
**Sources**  
[Why does my Bedrock Nova OCR still output “box” symbols instead of clean text in rebuilt PDFs? | AWS re:Post](https://repost.aws/questions/QURoHsjToiTE2nemqSg5QUIA/why-does-my-bedrock-nova-ocr-still-output-box-symbols-instead-of-clean-text-in-rebuilt-pd-fs)  
[Image understanding limitations - Amazon Nova](https://docs.aws.amazon.com/nova/latest/userguide/modalities-image-limitations.html)

Why does Amazon Nova OCR still produce “box” symbols (□/■■■) instead of real Chinese text in rebuilt PDFs?

Add your answer

Relevant content