ما تراه العين في الساعة الثالثة ليس ما كانت تراه في الأولى.
كشف وتصنيف وعدّ ومراقبة جودة: نموذج مدرَّب على صوركم لا يتعب، ويطبّق المعيار نفسه على كل قطعة، ويشير إلى ما يخرج عن المعيار في لحظة خروجه.
- مدرَّب على صوركم
- قابل للنشر دون اتصال
- عتبة تُحدَّد معكم
Visual inspection is the first thing to degrade
Not through carelessness: because human attention on a repetitive task is not constant, and nobody can make it constant.
The criterion drifts
Two operators do not classify the same borderline part identically, and the same operator does not classify it identically at the start and end of a shift. The standard exists on paper, not in practice.
Inspection is partial
Inspecting the whole flow costs too much, so you sample. The defect that slips between two samples is discovered by the customer.
Incidents are undocumented
When a problem comes back, there is no image, no timestamp, no trace. The analysis is done from memory, and the cause stays uncertain.
From image to decision
The hard part is not detecting: it is deciding where to place the threshold, and holding it.
Building the image set
We start from your real images, in your real lighting, framing and soiling conditions. Borderline cases matter as much as obvious ones: they determine the final quality, and they are usually missing from the sets we are given.
Annotation and defining the criterion
With your domain experts we fix what counts as a defect, a class, an object to count. This step almost always surfaces internal disagreement on specific cases — resolving it is part of the work, not a precondition.
Training and validation
The model is trained then evaluated on images it has never seen. Both errors are measured separately: the missed defect and the false alarm, which carry neither the same cost nor the same handling.
Setting the threshold
You decide the trade-off: miss fewer defects at the cost of more false alarms, or the reverse. This is a quantified business decision, not a technical parameter left at its default.
Deployment and monitoring
The model runs on site, without a connection if needed, on a machine or box near the camera. Uncertain cases are archived with their image and feed retraining: the system improves over time instead of drifting.
What the system can do
Object detection
Locate and count elements in an image or video stream, with their position and timestamp.
Classification
Sort images into categories defined with you: compliant, rework, scrap, or your own business classes.
Quality inspection
Detect scratches, missing parts, assembly faults, dimensional or appearance deviations on a line.
Code and plate reading
Barcodes, QR codes, serial numbers, licence plates, including degraded or partly obscured ones.
Counting and flow
Count passages, parts, vehicles or people, aggregated by period.
Visual document analysis
Cases where the image carries the information: drawings, schematics, field photos, incident reports.
Three environments, three constraints
End-of-line visual inspection is done by sampling. Customer returns concern appearance defects the sample missed.
In-line camera, model trained on the defects actually observed, threshold set with the quality manager, archiving of every uncertain case.
Inspection moves from sample to full flow, with an identical criterion at any hour and an image kept for every alert.
Serial numbers and codes are keyed in by hand at goods-in, with errors that propagate through stock.
Automatic reading at receipt, immediate check against the expected delivery note, alert on discrepancy.
Data entry disappears and discrepancies are caught at the dock, not at stocktake.
Counting of passages and occupancy is done by estimate, which allows no fine capacity planning.
Automated counting on the video stream, aggregated by period, with no identifying image retained.
Real figures replace estimates, with processing designed not to create personal data.
What we measure at pilot
Both errors are measured separately, on your images, in your conditions. A headline accuracy figure that does not separate them means nothing.
- of the flow inspected instead of a sample, on the covered station
- 100 %
of the flow inspected instead of a sample, on the covered station
- internet connection required: the model runs on site if you want it to
- 0
internet connection required: the model runs on site if you want it to
- between pilot launch and the decision to industrialise
- 6-10 weeks
between pilot launch and the decision to industrialise
هذه المقادير مستمدّة مما نقيسه في المرحلة التجريبية على نطاقات مماثلة. أما على رصيدكم فتُقاس قبل التصنيع، ولا يُوعَد بها مسبقًا.
أسئلة شائعة
It depends on how variable the defect is, not on an absolute number. A few hundred well-chosen examples including borderline cases beat ten thousand near-identical ones.
No. The model can run entirely on site, on a machine or box near the camera. That is the standard case in industrial environments.
Nothing, until you say how many defects are missed and how many false alarms are raised. We always give both figures separately.
We do not. Anonymous counting with no identifying image retained is possible; identifying people is not something we do.
It degrades if conditions change — new lighting, new product, new camera. That is why uncertain cases are archived: they feed periodic retraining.
Not automatically. We start from what exists and measure what it allows. When a camera is not good enough, we say so before the pilot, not after.
ما لا يفعله هذا الحل
Computer vision is not magic: on defects your own experts classify differently, no model will do better than they do. We turn down projects where the defect criterion cannot be defined, and we refuse to quote an accuracy figure without saying how many defects are missed and how many false alarms are raised. We also do not deploy person recognition: it is neither our trade nor ground we want to occupy.
Which visual inspection is slipping past you today?
Bring real images, borderline cases included. We will tell you frankly whether the criterion is learnable, or whether it needs defining first.