Bitread

Embedded OCR engine · patent pending

Full OCR on the chip, not in the cloud.

Bitread is a complete OCR engine for small processors. It finds the text in an image and reads it: labels, serial numbers, displays, forms and printed pages. Trained on your fonts, and 95% right on fonts it has never seen. A portable C library, a model under 1 MB, a text line in under a millisecond. No GPU, no network, no smart camera.

Test it on your images See the measured results
Nameplate · dot-matrixFrame 3
PASS SN G104-2610 0.25 ms per line · one desktop core

Measured on hard printed text

Our first benchmark uses the text that defeats general OCR: serial, date and part codes in four fonts: a 14-segment display, a dot-matrix font and two label fonts, each under five conditions: clean, scanned, low contrast, uneven light and display.

99.92%

characters read correctly: 16,922 of 16,935

20–54×

faster than Tesseract on the same lines and processor core

<1MB

read-only model, small enough for on-chip flash

74KB

library code compiled for a Cortex-M7, size-optimized

Results from our own tests. On the same lines, Tesseract made 8.0% errors on the dot-matrix code and 24.9% on the 14-segment display. Pilot partners get the test protocol and run it on their own images.

What it reads

Printed text, wherever the camera points.

Dot-matrix print

Text printed in dots by inkjet and impact printers, including faint, broken and touching characters that general OCR misreads.

Meters and displays

Seven- and fourteen-segment readouts on meters, scales and instruments, read on a battery-powered device instead of a photo sent to a server.

Labels and serials

Part numbers, lot codes and labels in ordinary print, with each reading checked against the format you expect before your device acts on it.

Barcodes and 2D codes

Linear barcodes and 2D codes are read by the same library as the text beside them, so one device reads the label, the serial number and its code in one pass.

Forms and pages

Whole pages, forms and receipts: Bitread finds the text blocks and lines, reads them in order and returns each word with its position, so the layout comes through with the text.

How it ships

A library you link, not a service you call.

Bitread is written in portable C and runs on the processor that already sits beside the image sensor. It reads text where the image is captured, so images never leave the device and a reading does not wait on a network.

Every reading carries a confidence. When text is uncertain, Bitread says so, so your device can read it again or reject it instead of passing a wrong code.

  • Platforms
    x86, 64-bit ARM and Cortex-M7, with identical output on each
  • Footprint
    74–146 KB of code, under 1 MB of model
  • Speed
    0.17–0.64 ms per text line on one desktop core
  • Conditions
    Clean, scanned, low contrast, uneven light and display images
  • Fonts
    Trained on your fonts; 95.08% of characters right on fonts never seen in training
  • Output
    Text in reading order, word positions, per-character confidence and format checks
  • Codes
    Linear barcodes and 2D codes, read alongside the text
  • Status
    Patent pending; pilots open. Handwriting (hand-printed forms) in development

Pilot program

Run Bitread on your own images.

Send us images of the text you need read. We report accuracy and speed side by side with your current reader, on your images and your target processor.

[email protected]

  1. Send 200 to 2,000 imagesFrom the camera you use now, under NDA.
  2. Get a side-by-side reportCharacters right, wrong reads, rejects and time per line.
  3. Pilot on your deviceOn your processor, running beside your current reader.