Industrial automation · Machine vision
Machine vision for industrial quality control
Conceptech builds machine vision systems for production lines in Israel: camera, lighting, compute and software that check every part, wired to your PLC and reject mechanism. It starts with a feasibility test on your parts.
Extensive experience servicing machines in aluminium plants and in food and beverage plants
- For
- Quality, operations and engineering managers at plants that inspect by eye or by sampling
- What you get
- An inspection station wired to your PLC, with automatic reject and a record for every part
- First step
- A feasibility test on good and defective parts from your line
What it checks
What can machine vision check on a production line?
Six common checks in automated and AI visual inspection. One station usually combines several of them, and the list for your product is set during the site visit.
Presence and absence
Is every component there: a screw, a seal, a cap, a label or a part in the assembly. The most common and simplest check.
Dimensions and position
Length, diameter, gap or angle measured from the image against a set tolerance. Needs stable lighting and calibration.
Surface defects
Scratches, cracks, stains, bubbles or paint defects. This is where deep learning helps most, because the defect looks different every time.
Labels, codes and OCR
Reading date codes, batch numbers and barcodes, and checking that the right label sits in the right place and orientation.
Counting and sorting
Counting items in a tray, a carton or on the conveyor, and sorting by type, colour or size.
Tracking on the conveyor
Following each item along the conveyor, so every result stays linked to the right part up to the reject point, even when parts are close together.
Choosing the technology
OpenCV, YOLO or a VLM: which technology fits?
When lighting is controlled and features are clear, classic OpenCV is enough. When the defect changes from part to part, you need deep learning. We test the choice on your parts, not on paper.
- Classic vision with OpenCV
- Suits measurement, presence and position when lighting is controlled and features are clear. Fast, predictable and easy to explain, with no large training set. Less suited to defects that change shape.
- Defect and object detection with YOLO
- A deep-learning model that finds defects, parts or missing components, even when they look different on every item. Needs a labelled image set from your line, and often a GPU-equipped computer at the line.
- Vision-language models (VLMs)
- Check an image against a written instruction, such as whether a label matches the order, or describe in words what the camera sees. Good for changing checks and documentation, less so for very fast lines.
- Object tracking
- Follows each item from frame to frame and along the conveyor, so every result belongs to the right part and you can count, sort and reject exactly what failed.
- A combination
- Often the answer is a mix: OpenCV for measurement, YOLO for defects and tracking to tie each result to its part. The mix is decided in feasibility, from the results on your parts.
Feasibility
How do you test feasibility before buying a system?
On your parts, before design and build. The decision to go ahead is made from results measured against targets agreed in writing up front.
A feasibility test answers one question: can the defects that matter to you be detected, at line speed, with the reliability you need? It comes before any station is designed, so the biggest risk in the project is tested while it is still cheap to test.
- Samples: you send good and defective parts, covering every defect type worth catching, and where possible images or video from the line.
- Lighting trials: we try several camera, lens and lighting set-ups until the defect shows clearly in the image. Much of the outcome is often decided here.
- A labelled set: we label the images together with your quality team, so the definition of a defect is yours.
- Written targets: before testing we agree targets, for example a detection rate per defect type and an allowed false-reject rate. These are targets to measure against, not a promise made in advance.
- Report: what was measured against each target, which technology fits, and what it takes to move to design.
When results fall short of the targets, the report explains why and what could change: lighting, camera angle, the defect definition or where the station sits on the line.
Line integration
Connecting to the line: PLC, automatic reject and data
A vision station is only worth something when the line acts on its result. That's why the PLC, reject and data connections are designed from the start.
- PLC: the station connects to the PLC already on the line, Siemens, LS Electric, Unitronics or Controllino, through digital I/O or over comms such as Modbus and EtherCAT.
- Trigger and tracking: a sensor or conveyor encoder tells the camera when to capture, and object tracking ties each result to its part up to the reject point.
- Automatic reject: a failed part leaves the line, for example by air jet, pusher or a line stop, depending on speed and product.
- Data: every result is stored with the image, time, line and shift. You see trends, and the data can be passed on to plant systems.
From there you can go further: an AI agent could turn a rejected part into a draft non-conformance report (NCR) with the image, line and shift, for the quality team to approve. AI agents for manufacturing
Machine vision is one part of our work on production lines. All industrial automation services
Process
What are the steps in a machine vision project?
Six phases, each with a clear deliverable and gate. The station is tested at our workshop and again at your plant, and we estimate each phase's duration after the initial evaluation.
01
Site visit
We see the line, the parts and the speed, and work out which defects really matter, which PLC is in place and where the camera and reject can go.
Output: List of inspections, requirements and constraints
Gate: You approve what is checked and what counts as a defect
02
Feasibility
Lighting and imaging trials on your samples, a labelled image set and measurement against the targets agreed up front.
Output: Feasibility report: results against targets, recommended technology
Gate: We decide together whether to move to design
03
Design
Camera, lens, lighting and compute selection, the enclosure and shielding, the interface to the PLC and reject mechanism, and the software architecture.
Output: Station design and acceptance-test protocol
Gate: Joint design review before parts are made
04 · FAT
Build and FAT
We build the station, tune the algorithm or train the model, and run a factory acceptance test at the workshop on an agreed set of parts.
Output: FAT results
Gate: The station passes FAT
05 · SAT
Installation and SAT
We install around your shifts, run the site acceptance test on the line and train operators and quality staff.
Output: SAT results, training and documentation
Gate: You sign off on the system
06
Support
Service on call-out or under a service plan, and model updates when a new product comes onto the line or a new defect type appears.
Output: Service plan and technical log
Machine vision for quality control: FAQ
How much does a machine vision system for a production line cost?
It depends mainly on what is inspected and on line speed, so the proposal is written after the site visit and feasibility test, not before. The main cost drivers:
- The number of cameras and views, and the number of defect types
- How subtle the defect is, and whether deep learning and a labelled set are needed
- Line speed, which sets the camera, lighting and compute
- The reject mechanism and the connection to the PLC and data systems
- The environment: dust, moisture, vibration or washdown
Feasibility is a separate step, so the decision on a larger investment is made after you've seen results on your own parts.
How accurate is machine vision, and what about false rejects?
There's no honest number before we've seen your parts, so we don't promise a percentage in advance. In feasibility we agree targets in writing, for example a detection rate per defect type and an allowed false-reject rate, and measure them on a labelled set from your line.
There is always a trade-off between the two: a system that catches every defect will also reject some good parts. The balance point is set with your quality team, weighing the cost of a defect reaching a customer against the cost of rejecting a good part.
Why does lighting matter so much in machine vision?
Because the camera only sees what the lighting shows: a defect that doesn't stand out in the image won't be caught, even by an advanced model. That's why feasibility tests several lighting set-ups, such as backlighting for silhouettes and dimensions, dome lighting for shiny surfaces or low-angle lighting for scratches. At the station, the camera is shielded from ambient light that changes during the day.
Can machine vision be added to an existing line?
Yes, and that's the common case. The camera and lighting mount on the existing conveyor or machine, a sensor or encoder triggers the capture, and the result goes to the existing PLC, even when another builder made the line.
During the site visit we check the physical space, the reject point and the PLC and comms already in place, and plan the installation around your shifts or a planned shutdown.
Why not just buy a Cognex or Keyence smart camera?
Sometimes that is exactly the right answer. Smart cameras from vendors like these are mature products with ready-made software and vendor support, and they suit well-defined checks such as presence, dimensions and code reading.
We aren't tied to any camera vendor, so we choose what fits each inspection: a smart camera, or an industrial camera with a computer running OpenCV or YOLO when you need flexibility, a custom model or a special connection to the line. The recommendation and the reasoning are in the feasibility report.
Will the system keep up with our line speed?
The required speed is defined at the site visit and tested in feasibility, not after installation. Line speed sets the exposure time, the camera type and the compute.
On a fast line we use strobe lighting and a global-shutter camera so the image doesn't blur, an encoder trigger so each capture happens at exactly the same point, and a GPU-equipped computer when a YOLO model has to run at the line. When a particular check can't keep up, we say so up front and propose an alternative.
Send us defective parts, and we'll see if the defect can be detected
Tell us what you inspect today, at what speed, and which defect reaches your customers. An engineer will reply with a first direction and a feasibility-test proposal.
Usually a reply within 3 working hours, by email or phone