Community

Hello,

I spent hours to test everything I could and what I found in web. However I still haven't resolved the issue why documents recognition performed through server doesn't include Field Extraction Training Batches. It's driving me crazy. Let me try to explain in more detail.

- I created a project using project type "Orders". In this project I use only one built-in document definition "Order".

- In the document definition Data Sets tab, in order to distinguish customers layouts I connected the database to the dataset "Variants".
- In the FlexiLaout tab I left selected "Use recognition settings from FlexiLaout".
- In the document definition settings tab (Document Definition Features) I have selected options "Training during processing" and "Enable clustering".
- In the properties of the fields that I want to capture I selected "Can have region" (General tab) and selected "Standard recognition" (Recognition tab)

- I didn't create batch type. I use the default.
- I didn't create classifier training batches because document definition "Order" is the only one that I use.
- In the Project Properties Workflow I have stages "Scanning", "Recognition", Verification", "Export" and "Training"

- In the Field Extraction Training Batches I loaded some documents, selected variant, selected field regions and trained (at least 5 documents per variant). Further documents and variants were added automatically when using the verification station.

The problem is that despite many trained samples, subsequent documents that automatically go to the verification station do not take into account the trained variants at all. Only selecting the "Clear Analysis Results" option and selecting "Recognize" again correctly finds the fields and values ​​in the document. It looks as if the server for some reason does not use "Fields extraction training" layouts.

It is also strange that if I add the document manually to the Working Batches List (via the Project Setup Station), "Recognizing documents" is performed correctly and all the necessary fields are filled in, and the logs show that the trained FlexiLayout variant was used (attached image)