Machine vision / Decision table
AOI vs AVI vs Manual QC: A Decision Table for E&E Manufacturers
Manual QC catches roughly 70 to 80 percent of defects under production conditions. AOI, automated optical inspection, catches above 95 percent of the defect classes it was programmed to find, because it measures geometry against a CAD reference. AVI, automated visual inspection, uses deep learning trained on your own parts to judge the appearance defects a rule cannot describe. The three are not competing products, they are gates for different defect families, and the wrong pairing is how a factory ends up paying for inspection twice. This article puts all three side by side across ten decision criteria, then maps them to the stations on a real line.
The numbers at a glance
01 / Three gates, not two
What each one actually does on the line
Most comparisons of automated optical inspection vs manual inspection stop at two columns, and that framing is why so many quality plans miss. There are three gates in common use on electrical and electronics lines in Malaysia, and each one owns a different family of defect.
The practical consequence: a factory that buys AOI to solve a cosmetic reject problem will get a machine that over-flags and gets switched off, and a factory that keeps three inspectors on a dimensional check is paying wages for a measurement a camera does better. The question is never which technology is best. It is which gate can physically see the defect that is currently escaping.
02 / The decision table
Ten criteria, three columns
This is the table to take into a capex meeting. Each row is a criterion that changes the answer, and the honest limitation is stated in the same cell as the strength.
| Criterion | Manual QC | AOI (rule / CAD based) | AVI (deep learning) |
|---|---|---|---|
| Detection rate | 70-80% under production conditions, best case 80% for a single inspector | 95%+ on programmed defect classes, 99%+ quoted on well-tuned solder and placement programmes | 97-99% on trained classes once the model has enough labelled examples of your parts |
| Consistency | Decays within the shift. Inspector-to-inspector agreement on severity runs 55 to 70 percent | Identical verdict on every unit, every shift, indefinitely | Identical verdict per model version. Changes only when the model is retrained |
| Defect family it owns | Contextual, cosmetic and one-off oddities. Anything requiring judgement against a customer's taste | Dimensional and positional: missing, wrong, reversed, offset, bridged, joint height and volume | Appearance: scratch, stain, dent, burr, discolouration, texture, contamination, weld and coating quality |
| Novel defects | Strongest here. A person notices something is wrong without being told what wrong looks like | Blind. A defect class absent from the programme is invisible at any accuracy setting | Partial. Anomaly-style models flag deviation from normal, but need review to classify it |
| False calls | Over-rejection rises with fatigue and with pressure after an escape | 30-80% of total calls on typical programmes, which is why a verification operator is part of the cost | Lower on appearance classes once trained, and it improves with every labelled review |
| Ramp-up | Days. Train to the standard and start | Weeks per product family. Every new product needs a programme written and validated, every revision an update | Weeks, driven by image collection. Needs enough real defect samples, which low-defect lines struggle to supply |
| Cost profile | Linear with volume and shift count. Around RM 1,900 to 2,500 base monthly per inspector in Malaysia, higher fully loaded | Capital first: roughly USD 3,000 for a 2D desktop unit to USD 50,000 to 200,000 for inline 3D, plus install, licence and maintenance | Capital plus engineering. Can often be retrofitted as an overlay on cameras or an existing AOI rather than a new machine |
| Traceability | None by default. A human decision leaves no image trail to audit | Image and measurement per unit, exportable to MES. Audit-ready evidence | Image, class and confidence score per unit, plus a defect trend history |
| Throughput | Capped by hands and eyes. Scales only by hiring | Line-rate, in-line, 100 percent inspection rather than sampling | Line-rate on GPU inference, subject to image capture speed |
| Best fit | Low volume, high mix, prototypes, and final judgement on flagged units | Stable high volume with measurable defects and a real customer PPM target | Surface and appearance defects, mixed part geometry, and defect classes no rule has captured |
Detection and false-call figures are published industry ranges, cited in the sections below. Price bands are indicative list ranges, not quotes. Salary figures are Malaysian base pay before statutory contributions, overtime and supervision.
03 / The human column
Why 80 percent is a ceiling, not a training gap
The number that should shape your inspection strategy is not a target, it is a limit. Research from Sandia National Labs found that a single human inspector detects about 80 percent of defects, and even two inspectors working in tandem top out near 96 percent. Broader studies of visual inspection under real production pacing put the working figure at 70 to 80 percent, with a wider band depending on defect type, lighting and cycle time.
What matters is the shape of that number. It is not a flat 80 percent across the shift, it is a curve that starts near its peak and decays. Inspection research shows accuracy dropping 15 to 25 percent after roughly two hours of continuous visual work, which is why recommended session lengths are 20 to 30 minutes. Agreement between two inspectors on defect severity runs only 55 to 70 percent, which means the same unit can pass on one shift and fail on the next without either inspector being wrong by their own standard.
Three further failure modes compound it. Resolution: sub-millimetre defects, hairline cracks and subtle colour shifts sit below the threshold of human perception at line speed. Shift effect: the same defect set is caught less often at night. And no record: a manual decision leaves no image trail, so when a customer sends back a failed unit there is nothing to audit and no way to find the systematic error behind it. We break this down further in why your inspectors catch four in five defects.
None of that is an argument for firing inspectors. It is an argument for not asking them to do the one job biology caps: staring at the same feature ten thousand times a shift and staying equally sharp on the last one.
04 / Pick by scenario
Match the gate to the situation
Criteria tables answer what each method is. This one answers what to do on Monday. Find the row that matches your line and read across.
| Your situation | Pick | Why |
|---|---|---|
| High volume, stable product, dimensional defects | AOI, in-line | Programming cost amortises across a long run, and 100 percent measurement replaces sampling at line rate |
| Cosmetic or surface rejects | AVI | Scratches, stains and texture cannot be expressed as a geometric threshold. A model trained on your parts can judge them |
| High mix, low volume, frequent revisions | Manual, plus AVI later | Programme writing and validation outruns the savings. Deep learning that generalises across similar parts closes the gap before rules do |
| Hidden joints under BGA, LGA or QFN | X-ray, not optical | No optical method sees through a package body. X-ray evaluates ball alignment, bridging and voids beneath it |
| Solder height, volume and coplanarity | 3D AOI | These defects are defined by height. A flat 2D image can infer them from shadow but cannot measure them |
| Customer demands per-unit evidence | AOI or AVI | Both produce an image and a measurement per serial number. Manual inspection produces a signature |
| New product ramp, defects still unknown | Manual first, then automate | You cannot programme or train for defect classes you have not seen yet. Collect them, then build the gate |
| Escapes reaching the customer now | Run both in parallel | Measure the escape rate of each gate on real production for a fortnight. The data settles the argument before capex |
Two rows deserve emphasis. The X-ray row exists because optical inspection is a line-of-sight method and pretending otherwise is how hidden-joint escapes happen. And the last row is the cheapest experiment in this article: parallel running costs a fortnight of operator time and it replaces every vendor claim, including the ones in this table, with your own numbers.
05 / The station map
Where each gate belongs in the line
Inspection strategy is a placement problem before it is a purchase problem. The same defect costs a different amount depending on where you catch it, so the cheapest gate is usually the earliest one that can see the defect at all.
| Station | Best gate | What it catches | Cost of missing it here |
|---|---|---|---|
| Incoming material | AVI or manual sampling | Wrong part, damaged lead frames, surface contamination, supplier drift | A bad part enters WIP and takes the whole assembly with it |
| Post-placement, pre-reflow | 2D AOI | Missing, misaligned, reversed or wrong component before solder sets | A ten-second fix becomes a rework station, a thermal cycle and a rework count against IPC-7711 |
| Post-reflow | 3D AOI | Joint height and volume, lifted leads, coplanarity, tombstoning, bridging | A latent open circuit that passes electrical test and fails in the field |
| Hidden joints | X-ray | Voids, ball alignment and bridging under BGA, LGA, CSP and some QFN | The most expensive escape class, because no downstream optical gate can find it |
| Final assembly | AVI | Cosmetic damage, wrong label, missing screw or gasket, cable routing, seal integrity | A customer-visible defect on a unit that is otherwise electrically perfect |
| Pack-out | AVI plus manual audit | Wrong count, wrong label, mixed variants, damage in handling | A logistics claim and a PPM hit for a defect that was never a manufacturing defect |
The pattern in that last column is the whole economics of inspection. A missing 0402 caught before the oven costs an operator ten seconds. The same part caught after reflow costs a rework station, a soldering iron, an unplanned thermal cycle, and a rework count against a customer specification that usually caps cycles per location at two or three. Caught by the customer, it costs a return, a containment sort, and a scorecard entry that follows you into the next quotation.
06 / The cost nobody tables
Automation does not remove the operator
Here is the line item that decides whether an AOI business case survives contact with the floor. AOI flags good units as defects, and reported false call ratios run anywhere from 30 to 80 percent of total calls depending on programme quality. Every one of those calls has to be judged by a person before the unit moves. So most lines station an operator at the machine, and that operator belongs in the AOI column of your cost model, not in the leftovers of the manual era.
The consequence is worse than the wage. The false call rate, not the detection rate, becomes the operational constraint on the system, because verification overhead erodes the throughput the machine was bought to deliver and introduces a second-order risk: genuine defects missed during fatigued manual review. An operator three hours into clicking through images of good solder is back at the 80 percent ceiling you bought the machine to escape.
This is where the AVI column earns its place in the comparison rather than sitting as a novelty. Deep-learning classification attacks both error modes at once, and documented retrofits show it: one contract manufacturer running automotive HDI boards took a false call rate that had crept from 12 percent to 30 percent and cut it to 3 percent with an AI overlay on existing AOI hardware, with the escape rate holding at zero. Read the full mechanics in our piece on reducing AOI false calls without letting defects escape, and the per-shift cost model in what a 1 percent versus 5 percent false call rate costs you per shift.
07 / On your parts
Trained on your defects, not a template
Every number in the tables above is an industry range. What decides your result is whether the gate is tuned to your parts, your tolerances, and your customer's acceptance standard. A generic recipe over-rejects on your line for the same reason a generic salary band tells you nothing about your payroll: the distribution that matters is local.
That is how CODETRACE machine vision and AOI is built. Models are trained on the customer's own good and defective parts, thresholds are set against your acceptance criteria, and JOVIS, our robotic vision inspection platform, handles micron-level geometry analysis so height-defined defects become a number your line can gate on rather than a judgement call. Systems are deployed and tuned on site across Selangor, the Klang Valley and Batu Kawan, which matters because the last 10 percent of accuracy always comes from tuning against the real board, the real lighting and the real conveyor.
CODETRACE is a member of the NVIDIA Inception programme, and the same engineers who scope the system commission it. If the honest answer for your line is that manual inspection is still the cheapest gate for the next twelve months, that is what we will tell you on site. That answer costs you nothing and it is worth more than a quotation for the wrong machine.
08 / Where to start
One afternoon settles most of this
You do not need a study to make this decision. You need four numbers and a defect. Write down the defect that is reaching your customer. Write down whether it is dimensional or appearance-based. Write down your monthly volume and your shift count. Then write down your current escape rate as PPM, even if it is an estimate from returns. Those four inputs eliminate most of the equipment catalogue in an afternoon.
Then run the parallel test. Keep your current process, add the candidate gate alongside it on real production, and compare what each one flags against what the customer finds. Two weeks of that produces a defensible escape rate for both gates, and it converts every claim in this article into a figure from your own floor. If the machine wins, you have a business case. If it does not, you have saved a capital request.
Inspection also sits inside a larger picture: it is one station alongside handling, packing and palletizing in factory automation, and the costs on both sides of the manual-versus-machine question are broken down in AOI vs manual inspection: where the costs actually sit. Get the gate right first. The rest of the line is easier to plan once the defect stops shipping.
Machines on the measurable checks. Models on the appearance checks. People on the judgement.
FAQ / AOI, AVI and manual QC
Questions, answered.
01What is the difference between AOI, AVI and manual inspection?
Manual QC is a trained person judging a part against a written standard, and it detects roughly 70 to 80 percent of defects under production conditions. AOI, automated optical inspection, is a camera and lighting system that compares a board or part against a CAD reference or golden sample using programmed rules, and detects above 95 percent on the defect classes it was programmed for. AVI, automated visual inspection, uses deep learning trained on your own good and bad parts instead of fixed rules, which lets it judge appearance-type defects such as scratches, stains, burrs and texture variation that a rule cannot describe. The three are gates for different defect families, not three brands of the same thing.
02Is automated optical inspection better than manual inspection?
On any repeatable, programmable check, yes. Published research puts human visual detection at 70 to 80 percent under real production conditions, with accuracy falling 15 to 25 percent after about two hours of continuous work, while AOI benchmarks sit above 95 percent and do not fatigue on the thousandth unit of the shift. Manual inspection still wins on three fronts: novel defects nobody programmed for, contextual or cosmetic judgement calls, and very high-mix low-volume work where programming time exceeds the savings. Most E and E factories in Malaysia end up running a hybrid line rather than replacing one with the other.
03When should I choose AVI instead of AOI?
Choose AVI when the defect cannot be written as a rule. If your reject reason is a scratch, a stain, a dent, a burr, a colour shift, a texture change, or an assembly that is wrong in a way that varies unit to unit, a deep-learning model trained on your own parts will outperform a geometric threshold. Choose AOI when the defect is dimensional or positional: missing component, wrong part, polarity, offset, solder bridging, joint height and volume. Many lines run both, with AOI on the measurable checks and AVI on the appearance checks that used to be the reason a human was standing there.
04Can automated inspection fully replace human inspectors?
No, and any vendor promising it should be treated carefully. Every automated inspection station still needs a person to re-judge flagged calls, because false call ratios on AOI programmes run anywhere from 30 to 80 percent of total calls depending on programme quality. People also remain the gate for unanticipated failure modes and for customer-specific acceptance judgement. What changes is the job: inspectors stop scanning every unit and start verifying calls and maintaining inspection programmes, which is more trainable and worth more on a payslip.
05How do E&E manufacturers in Malaysia decide between them?
Start from the defect that is reaching your customer, not from the equipment list. Write down the defect, the volume, the shift count, and whether the defect is dimensional or appearance-based, then match it to the gate that can physically see it. CODETRACE builds machine vision and AOI systems for semiconductor, electronics and automotive producers across Selangor, the Klang Valley and Batu Kawan, trains models on the customer's own parts rather than a generic library, and is a member of the NVIDIA Inception programme. The honest answer for many lines is a mixed set of gates, and we will say so on site.