DOCUMENT_TEXT_DETECTION API Japanese text recognition and old form of character inference

[OyaDuck]

Hello everyone,

We are using the DOCUMENT_TEXT_DETECTION API, an OCR service of Vision API. However, since around 9:00 AM (JST) on March 8, 2024, we have confirmed that some Japanese (JA) text recognized by the API includes old forms of Japanese characters.

For example, the modern character "内" (the standard in Japan) is now being returned as the old form character "內" in the inference results. We have also confirmed that other characters, such as "検" (standard) becoming "檢", are also being inferred as old forms of characters. It is highly likely that other characters will also be inferred as old forms of characters.

This problem has not occurred in the past, and we have confirmed that the same problem is still occurring after March 8, 2024.

We also checked the locale of the response, and at first we expected this problem to only occur when the locale is determined to be "und". However, we have confirmed that the same problem occurs even when the locale is determined to be "ja".

Has there been a change in the model or algorithm? Or is there a problem with the way we are using it?

If there is a solution, we would appreciate it if you could let us know.

UPDATE:

The situation has changed.

We changed the model of DOCUMENT_TEXT_DETECTION from "builtin/stable" to "builtin/weekly" and this problem was solved.

We think this problem is quite critical in Japanese OCR.
Does Google have any plans to reflect this to builtin/stable?

[Poala_Tenorio]

It sounds like you're encountering unexpected behavior with the Vision API's DOCUMENT_TEXT_DETECTION service, specifically related to Japanese (JA) text recognition. This could indeed be due to changes in the model or algorithm used by the API, or it could be a problem with your usage.

Here are some steps you can take to investigate and potentially resolve the issue:

1. Check for API Updates: Review the release notes or announcements from Google Cloud regarding updates or changes to the Vision API, especially around the time you started noticing this issue. There might be mentions of improvements or adjustments to the OCR model that could explain the behavior.

2. Review API Documentation: Double-check the documentation for the Vision API, paying close attention to any sections related to Japanese text recognition. Look for any changes in recommended usage or parameters that might impact the results you're seeing.

3. Test Different Configurations: Experiment with different configurations when making requests to the API. This could include adjusting parameters related to language detection, image preprocessing, or text output options. See if any changes influence the recognition results.

4. Contact Google Cloud Support: If you're still unable to resolve the issue, consider reaching out to Google Cloud support for assistance. They can provide more specific guidance based on your usage and may be able to escalate the issue to the appropriate team for investigation.

5. Consider Alternative Solutions: If the issue persists and you're unable to find a resolution, you might need to explore alternative OCR solutions or approaches for processing Japanese text. There are other OCR services and libraries available that you could try integrating into your workflow.

By following these steps, you should be able to gain a better understanding of the issue and hopefully find a solution or workaround.