Improving OCR Performance in Docling and Activation Conditions

[macgyvercsehdev]

Question

---

I am experiencing an issue with the processing time of OCR using Docling. I would like to better understand under what conditions OCR is activated and how I can significantly reduce the execution time.

Details of the Problem:

• Code Executed:

  time docling ./test01.pdf --pdf-backend dlparse_v4   --no-enrich-formula --force-ocr  --no-enrich-code  --ocr   --ocr-engine easyocr   --ocr-lang en   --device auto  --num-threads 8   --to md

• Execution Time:

  • real 49m43.488s
  • user 71m3.873s
  • sys 40m28.598s

• Test File:
  https://drive.google.com/file/d/1gX3say7OZ7danpJq9jAGIudf4zoFwYSN/view?usp=sharing

Questions and Requests:

1. OCR Activation Conditions: How can I determine when Docling will activate OCR? Are there specific settings that influence this decision?

2. Average Execution Time: What is the expected average execution time for OCR in Docling? The time obtained in my test seems excessive.

3. Processing Time Optimization: Are there ways to significantly reduce the processing time using cpu?

   • Parameter Optimization: Are there specific parameters that can be adjusted to improve performance?
   • System Configuration Improvement: Suggestions for hardware or configurations that can improve OCR performance.

Version of Docling:

docling 2.28.0
docling-core 2.23.2
docling-ibm-models 3.4.1
docling-parse 4.0.0

[cau-git]

To give a quick breakdown:

1. --ocr just enables OCR in general. It is only applied to bitmap areas detected in a given PDF. The digital PDF text cells are preserved and take precedence. There is a threshold parameter that you can set, which defines the minimum area of a page that must be covered with bitmap resources for OCR to engage: [pipeline_options.bitmap_area_threshold](https://github.com/docling-project/docling/blob/v2.5.2/docling/datamodel/pipeline_options.py#L26).
2. OCR confidence: Depending on the OCR model you choose, there are options to set a minimum confidence, such as here.
3. --force-ocr: This will override the default behaviour and always OCR the entire page, discarding any digital text cells. It is also the slowest option.

Additionally, there are significant differences in runtime performance between the OCR engines available, and most of them are slow on a CPU. You could try the RapidOCR on CPU as a best option, or ocrmac if you run on MacOS.