Scanning in a warehouse: native barcode detection, full-resolution OCR, and three rounds of correction
Choosing the browser's native BarcodeDetector and full-resolution still-photo OCR over a per-device commercial SDK licence, and what three rounds of correction taught — extract a vendor's service tag rather than the most confident string, do not crop away the pixels OCR needs, and show the operator which engine is actually running.
Status
Shipped, 11 July 2026.
The problem
Every device has to be identified by its serial or service tag. Typing them is slow and error-prone; the labels are small, often worn, and sometimes have no barcode at all — just printed text.
The decision: free, not paid
A commercial scanning SDK was considered and rejected. The system uses the
browser's native BarcodeDetector for barcodes and
full-resolution camera OCR for text-only labels.
The reasoning: a paid SDK is a per-device licence on hardware the business
keeps buying, for a job the platform can do. The trade-off accepted is that
BarcodeDetector support varies by browser, and OCR accuracy on a worn label
is worse than a dedicated engine's.
What the corrections taught
The first version worked in an office and poorly in a warehouse. Three rounds:
extract the Service Tag from vendor labels, not regulatory noise. A Dell label carries the service tag, the express service code, FCC identifiers, regulatory model numbers and certification marks. Naive OCR returned whichever string it read most confidently, which was frequently a regulatory number. The fix was to know what a vendor's tag actually looks like rather than take the clearest text.
** then — the scan box, added and then removed.** A tighter box that cropped detection was added to help the operator aim. It was then dropped in favour of full-frame detection and full-resolution still-photo OCR: cropping threw away the pixels OCR needed, and a live video frame is lower resolution than a still. Taking a photograph and reading that beat streaming.
resilient torch and on-screen engine status. The torch toggle failed on some devices and the operator could not tell which detection engine was running. Showing what the software is actually doing beats letting the operator guess why it is not working.
tag normalisation. The same device scanned from a barcode and read by OCR produced different strings. One normaliser, applied to both.
Why it matters
This is the point where a device enters the system. A misread tag creates a duplicate asset or attaches a wipe record to the wrong machine.
What was deliberately not built
No server-side OCR. The image never leaves the device, which keeps a photograph of a customer's asset label off the network.
Open questions
A dedicated USB barcode gun (Honeywell) was specified later and parked — see the scanner integration plan. The camera path is what is in use.
A published copy. Commit references and internal identifiers have been removed and the operator is not named; the engineering, the counts and the stated limits are unchanged.