Assigned
Status Update
No update yet.
Comments
ar...@google.com
ar...@google.com
Hi there,
I am unable to retrieve and test your image as the bucket you provided does not seem to be available anymore. Would it be possible for you to share the original image(twoc.png) that you’ve used in the Cloud Vision API Explorer?
Alternatively, it would be important to verify if any of the best practices would apply to your image. For example, when using TEXT_DETECTION the OCR of Cloud Vision API requires more resolution to detect characters and therefore, a minimum image size of 1024 x 768 pixels is advised.
I am unable to retrieve and test your image as the bucket you provided does not seem to be available anymore. Would it be possible for you to share the original image(twoc.png) that you’ve used in the Cloud Vision API Explorer?
Alternatively, it would be important to verify if any of the best practices would apply to your image. For example, when using TEXT_DETECTION the OCR of Cloud Vision API requires more resolution to detect characters and therefore, a minimum image size of 1024 x 768 pixels is advised.
Description
Issue summary: customers are facing a significant reduction in the quality of OCR (Optical Character Recognition) of documents containing Greek text starting early March 2024. Most importantly the issues included missing characters and misreading of Greek Δ characters as English A characters, affecting critical text required for indexing the documents. The issue is related to the promotion of Vision API model to built/stable on December 5, 2023 as it was not the case using builtin/legacy model. As a temporary workaround customer are pining for the builtin/legacy model until early June 2024, after which they will lose the model
Objective : Resolve regressions in OCR performance that are significantly impacting customers to use Vision API.
observed behavior : There are two types of issues customers are facing. In the first, characters are missing entirely from the results. In the second, Greek Δ characters reads as English A characters. Unfortunately this frequently happens in critical text that is required for indexing the documents.
Expected behavior : Include these characters in the results; read Δ characters correctly.
Future update: The builtin/stable has a big improvement on Latin datasets over the old model, but it seems it got this Greek character wrong for those images. builtin/weekly has an all-around improvement over builtin/stable, the ETA when we will swap it into the default version. It will be in Q3 this year. There are lots of evals run before builtin/weekly is updated to make sure the issue doesn't re-occur.