Welcome to AI-Powered OCR Solution

About This Project

This project — "Intelligently Extract Text & Data from Documents using OCR & NER" — was developed by Subrat Gupta as a practical implementation of modern Computer Vision and NLP techniques. The goal is to accurately extract key entities such as names, phone numbers, emails, and organizations from scanned documents, with a special focus on business cards. While this demonstration uses business cards for data privacy reasons, the framework is easily adaptable to documents like invoices, shipping bills, or other financial records.

To build this project, I integrated two core technologies in Data Science:

1. Computer Vision
2. Natural Language Processing (NLP)

In the Computer Vision module, the system scans the uploaded document, enhances its quality, and detects the position of textual elements. Then, using NLP techniques, it processes the extracted text, cleans it, and uses a custom-trained NER model to identify structured data.

Python Libraries used in Computer Vision Module:

• OpenCV – for image preprocessing and contour detection
• NumPy – for numerical operations
• Pytesseract – for Optical Character Recognition (OCR)

Python Libraries used in Natural Language Processing:

• SpaCy – for training and using the custom NER model
• Pandas – for data manipulation and structuring
• Regular Expressions – for pattern-based text cleaning
• String – for handling character-level processing

The entire project is divided into multiple development stages for better understanding and modular implementation:

Stage 1: Project Setup

• Install Python and necessary libraries
• Organize the file structure

Stage 2: Data Preparation

• Collect document images (e.g., business cards)
• Extract raw text using Pytesseract
• Clean and prepare text for annotation

Stage 3: Manual Data Labeling (BIO Tagging)

• Label entities using the BIO format:
  • B – Beginning of an entity
  • I – Inside of an entity
  • O – Outside or not part of any entity

Stage 4: Data Preprocessing

• Convert the labeled text into SpaCy-compatible training format
• Split data into training and testing sets

Stage 5: Model Training

• Configure and train the custom NER model using SpaCy
• Validate model accuracy on test samples

Stage 6: Deployment and Prediction

• Load and serve the trained model using Flask
• Render results with Displacy visualization
• Draw bounding boxes around predicted entities
• Build a complete document scanner web app

All of these stages come together in a fully functional web application built using Flask, where users can upload document images and get real-time predictions through the browser interface.

Developed By: Subrat Gupta
B.Tech CSE, GITAM University