Detecting Handwritten vs. Digital Docs in SAP CAP + DOX

[sundar]

I'm working on a solution using SAP CAP (Cloud Application Programming) with Node.js, integrated with SAP Document Information Extraction (DOX). This project sends documents via API to the DOX service to extract relevant information from files such as invoices and receipts. My requirement here is to check whether the document is handwritten or computer-generated. This distinction could significantly enhance data accuracy and processing logic. If anyone has insights on integrating handwriting detection with SAP DOX or using pre-processing techniques for document classification, I’d love to connect and learn more.

[daniel_dahlmeier]

Can you explain at what level the information handwritten vs printed text would be most helpful for your application? A the document level or field level?

[sundar]

hello @daniel_dahlmeier ,

Thanks for the question! For my use case, I only need to determine whether the entire document is handwritten or digitally printed. So, the detection is required at the document level, not the field level. The goal is to classify the document as either handwritten or digital to decide how it should be processed further in the workflow.

[daniel_dahlmeier · Accepted solution]

I don't think we expose a flag "handwritten" or "printed" at the moment. We can consider this feature in the future if there is more demand. In the meantime, you can explore using a vision LLM model (eg GPT4o, Gemini 2.0, Claude 3.7) from Generative AI Hub together with a custom prompt (eg "You are a helpful assistant for documents. Determine if the attached document contains handwritten text or only consists of printed text. Only return the word 'handwritten' or 'printed'. Do not return anything else")

https://learning.sap.com/learning-journeys/solving-your-business-problems-using-prompts-and-llms-in-...

[MartinSommer · Accepted solution]

Thanks 🙂 I am wondering how to retrieve a confidence rating or uncertainty indicator on the accuracy of the AI-generated output. In addition it would be interesting to know how you would handle a case where the model is not able to determine a meaningful output at all.

[daniel_dahlmeier · Accepted solution]

The Document Information Extraction service already returns field-level confidence values.

[daniel_dahlmeier · Accepted solution]

https://help.sap.com/docs/document-information-extraction/document-information-extraction/get-result...

[MartinSommer]

Maybe you can ask your question on one of the blog posts regarding OCR schema and handwriting detection 🙂

New Feature Release: OCR Only Schema - SAP Community

Handwriting Recognition by Document Information Ex... - SAP Community