OCR (optical character recognition) Turning a picture of a page into text a computer can search and check.
Short answer. OCR (optical character recognition) turns an image of text, such as a scanned page, a phone photo of an invoice or a PDF with no text layer, into machine-readable characters. A GST invoice shows why it matters: the supplier’s 15-character GSTIN, the invoice number and date, the HSN code and the tax amounts are printed fields that must be read exactly. Accuracy depends mostly on the scan: the open-source Tesseract engine works best at 300 dpi or more and loses accuracy on skewed, noisy or unevenly lit pages. Newer vision-language models can read pages without a separate OCR step, but the same rule applies: check the numbers against the original.
01
From pixels to characters
An OCR engine first cleans the image, then finds lines of text, then recognises the characters. Tesseract, the best-known open-source engine, added a neural-network recogniser focused on line recognition in version 4. It recognises more than 100 languages out of the box, outputs plain text, hOCR, PDF, TSV and other formats, and runs under the Apache 2.0 licence.
For a page with two scripts you load both languages: Tesseract’s documentation shows Hindi and English together as “-l eng+hin”, and notes that the order of the languages changes the time taken and the output.
02
A scanned GST invoice, field by field
The CBIC lists what a tax invoice must contain, and the GST Council describes the structure of the GSTIN. Use them as a checklist for what OCR output must be validated against. This is a summary for general information, not tax advice.
Field
What the sources say
What to check after OCR
Supplier’s GSTIN
15 characters: 2 for the State code, 10 for the PAN, 2 for the entity code, then a check character (GST Council)
Length, a valid State code, a PAN-shaped middle, the checksum; verify on the GST portal
Invoice number and date
A consecutive serial number, unique for the financial year, and the date of issue (CBIC)
Not misread; unique in your records
HSN code
HSN code of goods or accounting code of services, with the number of digits set by notification (CBIC)
All digits present, leading zeros kept
Taxable value, rate and tax
Taxable value after discount, the rate of tax and the amount of each tax charged (CBIC)
Taxable value times rate equals the tax shown; totals add up
Place of supply
Named with the State for inter-State supplies (CBIC)
Agrees with the State codes of both GSTINs and the tax type
03
What makes a scan hard to read
Factor
Why it hurts
What Tesseract’s documentation advises
Resolution
It works best at a DPI of at least 300
Rescan, or resize the image
Skew
Line segmentation quality falls significantly if the page is too skewed
Rotate so the text lines are horizontal
Noise and uneven lighting
Its automatic black-and-white conversion can be suboptimal on uneven backgrounds
Clean the image, or try another thresholding method; version 5 adds Adaptive Otsu and Sauvola
Dark scan borders
They can be read as extra characters
Crop to the text with a small margin
Light text on dark
Version 4 and later expect dark text on a light background
Invert the image
04
Three ways to read a page
Approach
How it works
Trade-off
Classic engine such as Tesseract
Image clean-up, line finding and a neural line recogniser, run locally
Needs a good scan; layout and meaning are left to later steps
Cloud document service
Digitises, extracts and classifies documents on the provider’s platform (Google Document AI)
Documents are processed in the provider’s cloud
OCR-free or vision-language model
One model reads the page image and outputs text or fields (Donut is one research example)
Its authors say it avoids OCR cost, inflexibility across languages and document types, and OCR errors passing on to later steps
05
Where ANVE stands
ANVE’s vision models read scanned documents, photos and diagrams, and its Bills function reads GST bills. Document question answering shows the source passage, so a person can check an answer against the scan. Models run on the box unless an administrator adds a cloud provider, and every model call is logged with whether it ran locally or in the cloud. Whichever tool you use, test it on a sample of your own invoices.
06
Questions people ask
Is OCR the same as AI reading a document?
OCR produces characters. Document AI and vision-language models go further, finding fields and tables and sometimes reading the page directly.
How do I improve OCR accuracy?
Tesseract’s documentation recommends scans of at least 300 dpi, straightening skewed pages, cleaning noise and cropping dark borders.
Can OCR read Hindi?
Tesseract supports Hindi and can load Hindi and English together. Accuracy depends on print quality and font, so test your own documents.
Does OCR keep my documents private?
It depends where it runs. Tesseract runs on your own machine, while a cloud service such as Google Document AI processes documents on the provider’s cloud platform.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
