On-device intelligence.
WHAT THE FRAMEWORKS PROVIDE
Running trained models on device
Vision analysis: faces, text, objects, barcodes
Natural language processing Speech recognition Sound classification
WHY ON-DEVICE MATTERS
Data does not leave the device It works offline There is no per-request cost Latency is low
WHY THAT MATTERS PARTICULARLY HERE
No connectivity requirement, and no data consumed.
WHAT TEXT RECOGNITION SUITS
Scanning documents, receipts, identity documents and codes.
WHAT TO KNOW ABOUT MODELS
They add to application size, sometimes substantially.
WHAT TO CONSIDER
Downloading models on demand rather than bundling them.
WHAT THE CONVERSION TOOLS PROVIDE
Turning models trained elsewhere into the platform's format.
WHAT TO TEST
Accuracy on data resembling your actual users' input.
WHY
Models trained elsewhere frequently perform differently on local data, including on local names, languages and documents.
WHAT TO NEVER ASSUME
That published accuracy figures apply to your case.
WHAT TO PROVIDE
A manual route when recognition fails.
