What defects can computer vision catch on an FMCG packaging line?

Vision systems on Indian FMCG lines reliably catch six defect families: unsealed or channel-leaking sachet seals, underfilled and overfilled packs, missing or cocked caps, unreadable or wrong date and batch codes, missing or skewed labels, and wrong-SKU packs after a changeover. Each one needs its own trained model and often its own camera angle and lighting.

Is computer vision better than a checkweigher for fill-level inspection?

No — they measure different things. A checkweigher measures mass and is the legally defensible instrument for net-content compliance. Vision measures visible fill height and catches faults a checkweigher misses, such as a correct-weight pack with foam or a crooked fill. Most lines that get this right run both.

What does an FMCG vision inspection system cost in India?

Cost is driven by defect-class count and line speed, not by the number of packs inspected. One camera watching one defect on a moderate-speed line is a different build from a multi-camera multi-SKU system at 120-plus items per minute. Industrial cameras, lighting and edge compute are quoted separately at supplier cost.

Can one vision system handle multiple SKUs without retraining?

Yes, if it is designed for it from the start. A model trained on variation across SKUs generalises to new variants; a model trained on one pack learns that pack. The practical constraint on an Indian FMCG line is changeover time, so recipe switching has to be an operator action taking seconds, not a retraining job.

Computer Vision for FMCG Quality Inspection in India (2026)

Q: How accurate is computer vision inspection at Indian FMCG line speeds?

A well-specified single-defect inspection reaches 99% or better at 120 items per minute. A Buteforce manufacturing QC deployment holds 99.2% classification accuracy at 120 items per minute and cut inspection errors 94% in its first month. Accuracy drops as defect classes are added, so per-class figures matter more than a headline number.

An unsealed sachet does not look like a problem on the line. It looks like a problem three weeks later, in a distributor's warehouse in Coimbatore, when a carton of spice powder has leaked into every pack around it and the whole case is a write-off. By then nobody can tell you which shift produced it.

That is the actual case for vision inspection on Indian FMCG lines, and it has nothing to do with AI being the future of manufacturing. It has to do with the fact that a human inspector watching 120 packs a minute is not inspecting. They are sampling, and hoping.

What defects can computer vision actually catch on an FMCG line?

Computer vision on an FMCG packaging line reliably catches six defect families: unsealed or channel-leaking sachet seals, underfilled and overfilled packs, missing or cocked caps and closures, unreadable or incorrect date and batch codes, missing or misaligned labels, and wrong-SKU packs that survive a changeover. Each family needs its own trained model, and usually its own camera position and lighting geometry. There is no single camera that watches a line and reports everything wrong with it — that product does not exist, and vendors implying otherwise are selling a demo.

The reason each defect needs its own treatment is physical, not algorithmic. A seal defect is a texture and edge problem on a reflective laminate, so it wants raking light from the side. A fill-level defect is a boundary problem inside a translucent container, so it wants backlighting. Put both on one overhead camera with one lamp and you get a system that detects neither well.

Seal and sachet integrity

This is where most Indian FMCG recalls start, and it is the defect worth solving first if you run sachet or pouch formats. A partial seal or a channel leak — a fold of product caught in the seal jaw leaving a microscopic path to the outside — is invisible in a top-down glance and obvious to a model trained on the seal band under angled illumination.

The hard part is not detection. It is the reject decision. On a vertical form-fill-seal machine running 60 to 120 packs a minute, the inspection point and the reject actuator are separated by a fixed number of packs, and if that offset is wrong you eject a good pack and pass the bad one. That timing is integration work with the PLC, and it is where projects that had a working model still failed.

Fill level, caps and closures

Fill level carries two separate costs. Underfill is a legal-metrology exposure under India's packaged commodities rules. Overfill is giveaway — product handed out free, permanently, at a rate nobody is measuring.

Vision is not a replacement for a checkweigher here, and any vendor who tells you it is has not run a line. A checkweigher measures mass and is the legally defensible instrument for net content. Vision measures what mass cannot see: a pack at correct weight with a foam head, a crooked fill, a container with the right total and the wrong distribution. Lines that get this right run both instruments and reconcile them.

Caps are simpler and higher value than they look. Missing and cocked-cap detection on oil, beverage and personal-care bottling is close to a solved problem for a trained model, and the cost of a missed one is a leaked case plus a retailer complaint.

Date codes, batch codes and labels

Date-code verification is the defect that ships thousands of units before anyone notices, because the printer does not fail loudly. It smears, drifts, runs low on ink, or prints a stale date after an operator forgot the changeover. A vision-OCR check reads the code the way a consumer would and compares it against what the batch record says it should be. Buteforce's document AI work sits on a dual-engine OCR architecture that returns a result in under a second — the same reading problem, in a harsher environment.

Label presence and alignment matters most where Indian FMCG actually differs from European FMCG: SKU velocity. A plant running eight variants of the same shampoo across a shift needs a model that generalises across pack variation rather than one trained on a single pack, and it needs changeover to be an operator dropdown, not a retraining request.

How accurate is computer vision inspection at Indian FMCG line speeds?

A well-specified single-defect inspection reaches 99% accuracy or better at 120 items per minute, and that is a realistic number to hold a vendor to. A Buteforce manufacturing quality-control deployment holds 99.2% classification accuracy at 120 items per minute and reduced inspection errors by 94% in its first month of operation. That system was built for an orthopaedic insole manufacturer, not an FMCG packaging line — the physics of the inspection transfer, the specific model does not, and pretending otherwise would be the kind of claim this industry is full of.

What matters more than the headline figure is the per-class breakdown. Accuracy degrades as defect classes are added to one model, and it degrades unevenly: seal detection might sit at 99.4% while a rare cap defect sits at 91% because there were only forty examples of it to train on. Ask any vendor for accuracy per defect class, with the false-positive rate alongside it. A system at 99.8% detection that stops the line four times a shift on good product will be switched off within a month, and that is the most common way these projects actually die — not with a wrong answer, but with a right answer nobody can afford to act on.

What are the real alternatives, and when is each one better?

Option	Typical fit	Where it wins over us
Cognex In-Sight / Dataman	Standard presence, barcode, OCR checks	Proven MTBF, global support, procurement-friendly. Days to deploy, not weeks.
Keyence vision sensors	Fast single-check installs	Fastest setup of anything on this list; strong local application-engineer support in India.
Optomech / Indus Vision	Indian machine-vision integrators	Deep PLC and line-integration experience; often already on your approved-vendor list.
Manual inspection + checkweigher	Low speed, low mix lines	Cheapest by far, and legally sufficient for net content. Genuinely correct below meaningful volume.
Buteforce custom vision	Defects no catalogue model was trained on	99.2% at 120 items/min on a trained custom model; you own the weights and the code outright.

The honest summary of that table: if your inspection is standard, buy a sensor. Cognex and Keyence have spent decades on reliability that a custom build does not match on day one, and their tooling exists precisely so you do not need a model trained from scratch. Custom work earns its price only when the defect is specific to your product, your laminate, your line — the case where a catalogue model has simply never seen what you need caught.

Not a fit if…

Do not start a vision project if you cannot show images of the defect you want caught. Without examples there is nothing to train on and no vendor can honestly commit to an accuracy number — anyone who does is guessing. Rare defects are the hard case here: if it happens twice a month, collecting a training set takes longer than the build.

Skip it too if the line runs under about 30 packs a minute with one SKU, where an inspector genuinely can watch every pack. If nobody on your side will own the system after handover, do not start at all; every abandoned system we have seen had no named owner on the plant side, and that predicts failure better than any technical factor. And if what you actually need is net-content compliance, buy the checkweigher first.

Where to start

Pick one defect — the one that has actually cost you a recall, a retailer penalty, or a batch write-off in the last year. Collect images of it for two weeks, good and bad, at line speed under production lighting. That dataset is worth more than any vendor evaluation, and it is the thing that turns a quote into a commitment.

We work on lines across the Chennai manufacturing corridor and publish what a vision inspection system costs rather than quoting on request. If you have the images, send us the defect — you will get an honest answer about whether it is worth building, including when it is not.