Industrial Vision Fits the Line Takt

Industrial vision turns factory line speed into a vision specification: fix the camera, light, and trigger first, then fit the whole decide chain inside the takt budget.

A 2000 ms line takt leaves a 300 ms vision slot, so fixed lighting plus a triggered camera must deliver a decidable image that a 190 ms chain can judge

Why Does This Exist?

A line that builds a part every $2$ seconds cannot pause for a human loupe. Defects cost recalls, missed screws cost returns, misread labels cost fines. Office vision assumes varied scenes and forgiving latency, while the factory floor fixes the scene and punishes delay.

Industrial vision is the line-aware umbrella over the standard skills. The decision skill is object detection, the reading skill is optical character recognition, and the sorting skill is image classification. What makes it industrial is not the model but the contract: a triggered image under controlled light in, a pass or fail decision with a logged reason out, inside the takt budget every cycle.

Use it when the camera, light, and part position are yours to fix. Use flexible scene models when the world refuses fixturing.

Think of It Like This

A bouncer with a checklist and a stopwatch

A club bouncer checks ID, stamp, and dress code in seconds per person, same three checks in the same order, with a lamp fixed over the door. Regulars get waved through fast because the pattern is known and the lighting never changes.

The encoder trigger is the knock on the door, the dome light is the fixed lamp, and the inspection recipe is the checklist. Where it stops: the bouncer can bend a rule for a familiar face, while a line station must hold its spec limits even when the shift supervisor pleads, because every exception is an unlogged defect escaping downstream.

How It Actually Works

Fix the world before training any model: mount the camera rigidly, light the feature with a dome or bar light so the defect always looks the same, and trigger on the encoder so each part lands on the same pixels. Budget time backwards from the takt: a $2000$ ms takt minus $1700$ ms of handling leaves a $300$ ms vision slot, and the chain of locate, then measure or classify, then decide against spec limits must fit inside it with margin.

Run locate first so measurement happens in part coordinates, not image coordinates, then compare against spec limits rather than vibes. Log every reject with its image crop so quality can audit the threshold weekly: the reject log is what turns arguments about sensitivity into arithmetic about false rejects versus escapes.

Worked example

A cap line runs $30$ parts per minute, one every $2000$ ms. Handling eats $1700$ ms, leaving $300$ ms for vision. The chain needs $120$ ms of model time plus $40$ ms of lighting settle plus $30$ ms of messaging, totaling $120 + 40 + 30 = 190$ ms, which fits inside $300$ ms with $110$ ms of margin.

On quality: $1000$ parts pass the station and the recipe flags $10$ good ones, so the false reject rate is $10 / 1000 = 0.01$ . On a defect-crop audit of $100$ crops where the recipe agrees with the auditor on $74$ , agreement is $74 / 100 = 0.74$ . The plant sees the $10$ scrapped good parts per thousand before any missed-defect math, which is why the threshold is priced in false rejects first.

Code

The point of this snippet is the budget check, not the model: it confirms the chain fits the slot and prices the false rejects from the worked example using only the standard library.

1takt, handling = 2000, 17002slot = takt - handling  # 300 ms for vision3chain = 120 + 40 + 30  # model + lighting + messaging4print(slot, chain, chain <= slot)5# -> 300 190 True6
7flagged_good, total = 10, 10008print(round(flagged_good / total, 3))9# -> 0.0110print(round(74 / 100, 2))11# -> 0.74

A $190$ ms chain inside a $300$ ms slot passes with $110$ ms of margin. A $0.01$ false reject rate scraps $10$ good parts per $1000$ , and the crop audit agrees at $0.74$ .

Watch Out For

Lighting drift that moves the decision boundary

A new skylight or a dusty dome shifts brightness by $15$ levels across a week, and the model reads that shift as dents. The symptom is a reject rate that creeps with the time of day while the parts stay identical. Fix it with enclosed lighting, a reference tile check each shift, and brightness-normalized features.

Lab accuracy that ignores takt and handling

A $99$ percent accurate model at $900$ ms still fails a $300$ ms slot and backs the line up. The symptom is perfect offline scores plus line stops on day one. Fix it by speccing latency and false-reject cost first, then picking the model that fits the slot.

The Quick Version

Industrial vision decides pass or fail inside the line takt, not at lab latency.
Fixed camera plus controlled light plus encoder trigger come before any model.
The fixture fits a $190$ ms chain inside a $300$ ms vision slot with $110$ ms of margin.
A $0.01$ false reject rate scraps $10$ good parts per $1000$ ; the crop audit agrees at $0.74$ .
Enclose the light and audit rejects weekly, or daylight rewrites the threshold.