Tech note: barcode and character recognition — AI, trained fonts or smart cameras?
Inkjet dates, laser marks, label barcodes and wire-sleeve numbers each have different pitfalls. Across four projects, 2D codes were almost always readable, while the smallest 1D barcode read only about 60% of the time.
Dates, lot numbers and barcodes on products are the backbone of traceability. But “readable” and “every one readable” are very different. Here is what Shin Shou learned in pharmaceutical, memory, label and control-panel projects.
Three ways to recognise characters
- Deep-learning OCR: a pre-trained model reads alphanumerics and some symbols without extra training and performs well; the drawback is longer computation, which must be checked against line speed
- Trained-font matching: characters are learned first, then matched; fast, and dot-matrix characters need not be connected; the drawback is that each font size must be learned separately
- Commercial smart cameras: well integrated and quick to deploy, but fonts still need training
Key test results
- Laser-marked barcodes (2 × 2 mm): read reliably at 20 μm and 30 μm with ring or dome light
- Laser-marked text (1 × 0.4 mm): regular fonts are read directly by AI; unusual fonts (some “8”s) are misread without retraining
- Label 2D codes: all readable
- Label 1D barcodes: most readable, but the smallest type read only about 60% of the time — enlarging it is recommended
- With incomplete OCR training, poorly printed characters may be read as other characters — a risk that needs a control step
Common issues with wire-sleeve numbers
Sleeve numbers on control panels read best when facing the camera; 0, O and Q are easily confused, and twisted or sideways numbers cause errors. Print larger with wider spacing, and build a position-to-part list so sequences can be checked.
Advice
Confirm font types, minimum character height and line speed before choosing a method. Enlarging barcodes or widening spacing at the design stage often helps more than a more expensive camera.
