document understanding and OCR

This project line groups my earlier public document-understanding work: LayoutLM classification, driver-license OCR/object detection, invoice field extraction, and SaastIE-style information extraction. It is the right place for the older Medium posts because the posts explain the build path behind the repositories.

How the projects fit together

The common thread is that document understanding is rarely just text classification. A useful pipeline has to decide what kind of document it is, where the relevant fields are, what text is attached to those fields, and how a person can validate the result. The repos in this group each cover one slice of that pipeline.

That is also why this page links the Medium posts rather than hiding them. The posts show the earlier implementation path: how I framed the document task, what model family I used, and where OCR or object detection entered the pipeline.

public work

• Document Classification using LayoutLM

  Layout-aware classification | PyTorch | Hugging Face Transformers

  Document image classification prototype using OCR words, bounding boxes, and LayoutLM sequence classification.

  github medium
• US Driver License Data Extraction

  YOLOv5 | object detection | field localization

  Driver-license extraction prototype for locating fields such as name, address, license number, gender, and date of birth.

  github medium
• Invoice Data Extraction

  OCR | field extraction | semi-structured documents

  Earlier invoice extraction repository for parsing useful fields from invoice-like documents.

  github
• SaastIE

  Information extraction | structured output

  Older information-extraction repository kept as part of the same public document-understanding trail.

  github

related writing

• Document Classification using LayoutLM · Medium, Apr 8, 2022
• Driver's License Data extraction using CNN (YOLOv5). Part-1 · Analytics Vidhya on Medium, Mar 3, 2021

What held up and what I would change now

The part that held up is the pipeline framing: documents need layout, text, visual region detection, and validation hooks. The part I would change today is the evaluation and packaging. I would separate demo notebooks from reusable library code, add clearer dataset/provenance notes, and make the review step explicit instead of treating extraction as the end of the workflow.