README.md

metadata

title: KYC OCR (India)
emoji: 📄
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.0.0
python_version: '3.12'
app_file: app.py
pinned: false
license: mit
short_description: Extract fields from Indian PAN cards and cheques on CPU

🇮🇳 KYC Document OCR

A lightweight, CPU-friendly pipeline for extracting structured fields from Indian KYC documents (PAN card, cheque) from images or PDFs.

Built on RapidOCR — ONNX-runtime deployment of PaddleOCR PP-OCRv4 models. No GPU needed.

What it does

• PDF or image input (PyMuPDF rasterises PDFs at 300 DPI)
• Document type auto-detection via keyword + regex scoring
• Cheque rotation correction — detects sideways-scanned cheques (text spans more vertically than horizontally) and rotates before MICR extraction
• Format-spec validation — PAN regex with 4th-character entity-type lookup, IFSC regex, date normalisation, MICR positional parse

Supported documents

Document	Fields extracted
PAN card	PAN number, name, father's name (individual) / company name, DOB / date of incorporation, entity type (Individual / Company / HUF / Firm / Trust / etc.)
Cheque	Bank name, branch address, IFSC code, account number, 9-digit MICR code, cheque number, transaction code
Aadhaar	Coming soon — recommended path: read the secure QR code via pyzbar rather than OCR
Passport	Coming soon — recommended path: parse the MRZ via passporteye

How to use the Space

1. Drop a PDF or image into the upload box
2. Click Extract fields
3. View the structured output and raw OCR (for debugging)

The included example is a synthetic PAN with fake data so you can try it without uploading real documents.

Running locally

git clone <this-space-url>
cd <repo>
pip install -r requirements.txt
python app.py

Then open http://127.0.0.1:7860.

Using kyc_extract.py directly as a library

from kyc_extract import process, RapidOCR

engine = RapidOCR()
results = process("path/to/document.pdf", engine)

for page in results:
    print(page["detected_type"])     # 'pan' or 'cheque'
    print(page["fields"])             # structured dict

Privacy

This is a public Space. Do not upload real personal documents. For production KYC pipelines, fork this repo and deploy on your own private infrastructure (DPDP Act and UIDAI guidelines require it for Aadhaar in particular).

Architecture notes

• All extraction is rule-based — no model is hardcoded to specific names, banks, PANs, or addresses. The same script handles any PAN or any bank's cheque.
• Document classification uses keyword + regex scoring; if confidence is low, type is unknown and no extractor runs.
• For cheques in portrait-orientation PDFs (common when phones are held vertically), the pipeline detects the layout anomaly and re-OCRs the rotated image, merging text from both passes so address and IFSC don't get lost in the rotation.

Limitations

• English-only OCR — Hindi label text (नाम/Name) comes out garbled but the pipeline has shape-based fallbacks so this doesn't break field extraction. For full Hindi extraction, add lang='hi' to RapidOCR.
• MICR fonts are E-13B and not perfectly read by general OCR. The pipeline reconstructs the 9-digit MICR from OCR fragments via positional parsing, which works on most cheques but degrades on poor scans.
• Aadhaar and passport extractors are stubs. The proper approach uses QR and MRZ respectively, not pure OCR.