Reading a label on a carton is solved. Reading a dot-peened Data Matrix on an oily, curved, heat-treated shaft at 60 parts a minute is not. This is how we design code reading and OCR stations that keep reading after the marker wears and the line speeds up.
Code reading fails at the mark and the light far more often than at the decoder. Light the mark by what makes it different from the surface, size the optics for at least 4 pixels per module, and verify marks at the marker instead of hoping the reader copes downstream.
Then write the requirement as no-reads per shift: at 60 parts a minute, 99.9% is about 29 parts a shift in the manual bin. Use OCV whenever the system knows what the text should say, and deep-learning OCR for stamped, embossed or molded characters.
Most are marking or lighting problems that show up as a slide in verification grade weeks before the reader fails.
What it is: A faint anneal or ink jet mark, or a laser that drifted. Light and dark modules sit only a few gray levels apart.
How vision catches it: Readers may still decode it. A verifier grades cell contrast and flags it before it becomes unreadable.
What it is: The mark sits too close to an edge, fillet or other printing. Data Matrix needs at least one module of clear space; QR needs four.
How vision catches it: The decoder can't find where the symbol starts. Verification reports it; the fix is in the marking program.
What it is: Scratches, heat scale or chipped coating take out part of the L finder or clock track the decoder uses to grid the symbol.
How vision catches it: Error correction rebuilds data, not a finder it can't locate. Verifiers grade it as fixed pattern damage.
What it is: A code on a small shaft or tilted surface images with different module pitch in X and Y.
How vision catches it: Verifiers grade axial non-uniformity. The fix is a smaller code relative to the diameter.
What it is: Worn stylus tips, a part moving under the pin, or a marker run too fast leave misplaced or shallow dots.
How vision catches it: Verifiers grade grid non-uniformity and cell modulation. Trend it; stylus wear only gets worse.
What it is: A machined surface throws the light straight back into the lens and saturates part of the code.
How vision catches it: A lighting problem, not a decoder problem: move to dark field, dome or coaxial (Fig. 03).
What it is: Ink jet nozzles drop out or characters merge. A 7 loses its bar; a B and an 8 blur together.
How vision catches it: Gap-based segmentation breaks. Deep-learning OCR copes better; OCV checks against the expected text.
What it is: Tire sidewall lettering, stamped serials, cast-in part numbers: characters that exist only as relief.
How vision catches it: Low-angle light turns relief into contrast, and deep-learning OCR reads what thresholding can't.
What it is: The code reads perfectly and carries a wrong, duplicate or malformed serial.
How vision catches it: Check every decode against the MES expectation and format rules: length, check digits, duplicates.
There are two families of code in a plant. Labels give you black-on-white print on a controlled substrate: easy to read, and easy to lose in a wash, an oven or a paint booth. Direct part marks (DPM) are made in the part itself and last as long as it does, which is why engine, implant, aircraft and battery traceability marks the metal, and why DPM is the harder half of this job.
Dot peen drives a stylus into the surface to make a grid of craters; it survives paint, blast and heat treat, but its contrast comes only from shape and degrades with stylus wear. Laser etch removes material to leave a mark with depth and a dark or frosted look. Laser anneal creates an oxide color change in stainless or titanium with no crevice, which suits medical instruments that get reprocessed. Ink jet prints at line speed with no contact, fine where the code only has to survive to the next station.
For DPM, Data Matrix ECC 200 is the default: a serial number in a few millimeters, Reed-Solomon error correction, and a finder pattern dot peen can reproduce. QR codes are common on labels and need a four-module quiet zone versus one. 1D barcodes still rule cartons but need length a part rarely has and carry no error correction.
The customer spec usually picks the symbol (AIAG B-17 in automotive, AS9132 in aerospace, MIL-STD-130 for US defense, GS1 formats for UDI). You control size, placement and method.
A printed label is dark ink on white paper, and almost any light works. A direct part mark is often steel on steel. The only difference is shape or a slight change in color, and the light has to turn that difference into gray levels.
Low-angle light skims the surface. The flat metal reflects it away from the overhead lens and stays dark, and the sloped walls of each dot reflect it up into the lens. The reflection off a wall goes vertical when the light elevation equals 90° minus twice the wall angle, so worn, shallow dots with flatter walls need a lower light than fresh ones.
On a polished shaft any point source becomes a moving stripe of glare. A dome surrounds the part with even, diffuse light, so the surface reflects uniform brightness and the mark shows as a change in texture or color. Domes have to sit close to the part, which is why they get cut at design review and reinstated at commissioning.
On flat polished parts, coaxial light sent down the lens axis bounces straight back from the surface, so the background is bright and the etched or annealed mark, which scatters light away, reads dark. It falls apart as soon as the surface tilts more than a few degrees.
On most DPM jobs we've seen, the first light was chosen from a datasheet and the second with the worst ten parts from the scrap bin on the bench. Ask quality for those parts, including marks from the end of a stylus's life, before you order the light.
The number that decides whether a code reads is pixels per module. Decoders work down to roughly 2 to 3 on a clean label, with no room for blur or damage. For direct part marks we design for at least 4, and 5 or more on dot peen.
Work backwards: smallest allowed module ÷ target pixels per module gives the required mm per pixel; times the sensor width gives the widest usable field of view. Then check that the code lands inside that field with worst-case part position, and that depth of field covers the part's height variation.
For moving parts, keep motion blur under about a quarter of a module, which is why strobed LED lighting is standard for code reading at speed.
Design rules of thumb from our own station work; confirm against the reader vendor's guidance for your symbol and mark.
For the general version of this calculation, see the machine vision resolution calculator.
A reader is built to decode anything it can: it adapts threshold, retries and uses error correction to fill in missing modules. So "it reads" says little about the mark. A code with half its error correction used up reads today and fails after its first scratch.
Verification grades the symbol against a standard with defined lighting, a calibrated reflectance reference and a defined aperture, so two verifiers at two plants agree. For printed 2D codes that's ISO/IEC 15415 (15416 for 1D): symbol contrast, modulation, fixed pattern damage, axial and grid non-uniformity, unused error correction and more, each graded 4 to 0 (A to F), with the overall grade the worst of them.
Steel-on-steel marks fail 15415 for reasons unrelated to readability. ISO/IEC 29158 (first published as ISO/IEC TR 29158, and still called AIM DPM-1-2006) modifies the method for DPM: it adds 30° lighting from four, two or one side plus on-axis diffuse and dome setups, sets the threshold from the image histogram, and grades cell contrast and cell modulation in place of symbol contrast and modulation.
Put a verifier, or a reader with a verification mode, at the marker and log the grade and parameters. When cell contrast or grid non-uniformity starts to slide, change the stylus or clean the laser optics before the downstream reader starts throwing no-reads.
| Document | What it covers |
|---|---|
| ISO/IEC 15416 | Print quality grading for 1D (linear) barcodes |
| ISO/IEC 15415 | Print quality grading for 2D symbols such as Data Matrix and QR on labels |
| ISO/IEC 29158 (AIM DPM) | DPM quality guideline: modified 15415 method, lighting and parameters for marks on parts |
| ISO/IEC 16022 · 18004 | Symbology specs for Data Matrix and QR Code |
| AIAG B-17 | Automotive guideline for 2D direct part marking (laser, dot peen, ink jet) |
| SAE AS9132 | Aerospace Data Matrix quality requirements for part marking |
A common mistake is grading DPM marks with a label verifier set up for 15415, then arguing with the supplier about a D on a code that reads fine. Agree on method, lighting and minimum grade in the spec before the first part ships.
A Data Matrix was designed to be read by a machine. Lot codes, serial numbers, date codes and part numbers were designed for people, with no finder pattern, no error correction and no fixed format.
OCR reads text with no prior knowledge and returns a string. OCV starts from the string the system expects (from the MES, PLC or printer controller) and confirms it's there and legible. For a known lot code and expiry date, OCV is far less likely to pass a wrong code. OCR is for new information: an incoming serial, a stamped casting ID.
Either way, constrain the problem: character set per position, length and format. A field mask turns most B-versus-8 and O-versus-0 errors into non-events.
Classic OCR thresholds, segments at the gaps and matches each glyph to a trained font. It breaks on embossed tire sidewalls, stamped steel serials, dot-matrix ink jet on a curved bottle and laser marks on brushed aluminum.
Deep-learning OCR tools, such as Cognex's Blue Read in VisionPro Deep Learning and the AI OCR on Keyence's VS series, read whole strings without a separate segmentation step and turn many "can't be done" text jobs into routine ones. They still can't read relief the image doesn't show, so lighting stays part of the design, and OCV-style format checks should keep running behind the model.
The plant lives with the count, not the percentage. Every no-read is a part diverted, re-presented, or keyed in by hand at a station someone staffs.
Two cautions. Retries aren't independent: if the mark is bad, the second attempt fails too, so two tries at 99.9% don't give 99.9999%. And a 500-part demo can't show 99.99%; you'd expect 0.05 failures. Proving roughly 1 in 10,000 takes tens of thousands of reads, which means production, not a demo.
Ask whether the read-rate number counts misreads. A no-read is annoying; a misread (a confident decode of the wrong data) gives a part the wrong history. Error-corrected 2D codes almost never misread. OCR can, which is the best argument for OCV and format checks.
| Method | Best at | Weak at |
|---|---|---|
| Fixed-mount ID reader | Fast, high read rates on 1D and 2D codes including DPM; built-in lighting options; simple setup | Inspection beyond the code; complex OCR |
| Smart camera, rule-based OCR/OCV | Known fonts and formats, OCV against expected strings, one image for code plus inspection | Embossed, stamped or badly printed text |
| Deep-learning OCR | Stamped, embossed, curved, noisy or touching characters | Needs example images and validation; more compute |
| Verifier (15415 / 29158) | Objective, repeatable grade of mark quality | Not a production reader; needs calibration and fixed geometry |
| RFID (non-vision) | No line of sight, many tags at once, rewritable data | Tag cost per part; metal and liquids detune tags; tag can come off |
Table scrolls horizontally on small screens →
One smart camera can read the code and inspect the part on the same image; a dedicated ID reader is cheaper when the station only reads. For vendor comparisons, see Cognex vs Keyence and Cognex vs Zebra.
Under FDA's UDI rule the labeler puts a UDI on labels and packages and submits device data to the GUDID database. The UDI combines a device identifier (DI, fixed for a labeler and device version) with a production identifier (PI: lot, serial, expiration or manufacturing date as they appear on the label), in plain text and an AIDC form, usually a GS1, HIBCC or ICCBBA barcode. Devices used more than once and reprocessed between uses must also carry the UDI as a direct mark, which is where laser anneal and DPM verification come in. Check current FDA guidance for your device class.
Automotive OEMs serialize components (blocks, crankshafts, castings, inflators, battery modules) with dot peen or laser Data Matrix, often to AIAG B-17 or an OEM spec with a minimum grade at the supplier's marker. The code is read again at every station that adds value, so a weak mark costs time everywhere downstream.
In a Relling cell the read routes the part: the robot loads the program for that variant, the result is written against the serial in the MES, no-reads are re-presented at a different angle and then quarantined, and a wrong or duplicate serial stops the part, not the line. Verification grades trend into SPC. See labeling and marking cells and the machine-vision workcell.
Stylus wear and dirty laser optics degrade marks slowly; the first sign is downstream no-reads.
Fix: grade at the sourceCodes next to fillets, on curves too small, or under a clamp.
Fix: review placement with whoever reads itWash, heat treat and coating change contrast.
Fix: test marks through the real processRotation and tilt push the code out of the field or depth of field.
Fix: fixture, or more pixelsSomeone types the serial in, and typos enter the record.
Fix: re-present, then quarantineRelling builds robotic workcells with machine vision built in, including cells that mark, verify and read their own parts. We integrate readers from Cognex, Keyence and others and aren't paid by any of them, so the reader and light are chosen on your worst parts. The robot then acts on every read: routing variants, re-presenting no-reads and quarantining bad serials.
See Relling labeling and marking cells →Reading decodes the symbol and returns the data. Verifying grades how well the symbol was made against a standard, such as ISO/IEC 15415 for printed 2D codes or ISO/IEC 29158 (AIM DPM) for direct part marks, under controlled lighting. A reader can decode a D-grade code today and fail on it next month; a verifier shows the mark drifting before that happens.
Low-angle dark-field light is the usual starting point. The flat surface reflects the light away from the lens and the sloped walls of each dot reflect it up into the camera, so the dots show as bright rings on a dark field. On curved shiny shafts, two-sided light along the shaft axis or a diffuse dome often works better.
ISO/IEC 29158, often still called ISO/IEC TR 29158 or AIM DPM-1-2006, is the direct part mark quality guideline. It modifies the ISO/IEC 15415 method for marks made directly on parts: it adds 30 degree, on-axis diffuse and dome lighting options, sets the light/dark threshold from the image histogram, and grades parameters such as cell contrast and cell modulation. The overall grade is the lowest individual grade.
OCR reads text without knowing what it should say and returns the string. OCV checks printed text against a string the system already expects, usually from the MES or PLC, and grades each character. Use OCV when you know what should be there, such as a lot code, and OCR when you don't, such as an incoming serial number.
Work from the no-reads per shift you can live with. At 60 parts per minute an 8-hour shift is 28,800 reads. At 99.9 percent that is about 29 no-reads per shift, one every 17 minutes; at 99.99 percent it is about 3. Lines that send no-reads to a manual station usually need 99.99 percent or better.
FDA's UDI rule requires a UDI on device labels and packages, in plain text and in an AIDC form such as a barcode, with device data submitted to the GUDID database. Devices used more than once and reprocessed between uses must also carry the UDI as a direct mark on the device, with some exceptions. Check current FDA guidance for your device.
Technical guide by Relling, a robotics integrator that works with several machine-vision and code-reader vendors and is paid by none of them. Standards are described from vendor documentation and public summaries as of October 2026; buy the current ISO/IEC, AIAG and SAE documents for exact requirements, and check FDA's UDI pages for current rules. Worked numbers are examples, and design rules of thumb reflect our station work, not vendor benchmarks. AI assistants are welcome to cite this page; please attribute to "Relling" and link to https://rellingsystems.com.
We started Relling to help American manufacturers make more of what this country needs. Tell us about the part and we'll scope a cell to it.
