Let the line catch the defect — not the customer.
Manual inspection is slow, inconsistent, and tired by the end of the shift. CISH designs, builds and commissions automated machine vision inspection stations for production lines in South Africa — industrial cameras, engineered lighting, and inspection algorithms we develop ourselves on the HIKROBOT VisionMaster platform — to detect defects, verify labels and codes, check fill level and caps, measure dimensions, and reject bad product at line speed, on 100% of production.
Machine vision, in plain terms.
A machine vision inspection station is a camera (or several), purpose-chosen lighting, and software that analyses every product as it passes — in milliseconds. Where a human inspector samples, tires, and disagrees with the next inspector, a vision system checks 100% of production against the same criteria, logs every result, and triggers a reject when a product fails. The same station can also read and verify codes, measure parts without contact, count items, and hand the results to your OEE or quality system.
Three terms get mixed up in buyer conversations. Machine vision is the industrial discipline: cameras, lighting, optics and software engineered for a specific check on a specific line. Automated optical inspection (AOI) is the same thing described from the quality side — a vision system doing the inspection a person used to do. Computer vision is the broader field of software that interprets images, from phone apps to autonomous vehicles; machine vision is the part of it that has to work at line speed, in a dusty plant, for years. A vision sensor is the simplest form: one camera with a fixed job, such as "is the cap on?".
The CISH difference: we don't resell cameras. We design the inspection algorithm for your specific product and defect, build the station, integrate it with your line's PLC and reject mechanism, and tune it on your real product until the false-reject and missed-defect rates are where they need to be.
The checks we build most often.
Defect & surface detection
Cracks, scratches, contamination, flash, short-shots, voids, colour deviation, and surface blemishes on moulded parts, packaging, castings, and finished goods. Deep-learning models handle defects that are too variable to describe with rules.
Dimensional measurement
Non-contact measurement of length, diameter, gap, angle, and position against tolerance — at line speed, on 100% of product, not a sample. Calibrated to engineering units, with 3D laser profiling where height or volume matters.
Code & date verification (OCR/OCV)
Read and verify printed date codes, lot and batch numbers, and 1D/2D barcodes — catching missing, illegible, or wrong codes before product ships. OCR reads the characters; OCV verifies they match what the batch should say.
Label & print inspection
A label inspection system checks label presence, position, skew, wrinkle, correct artwork and SKU (mix-up detection), and print quality — critical for retail compliance, the foodstuffs labelling regulations, and brand consistency.
Fill level, cap & seal inspection
Fill level inspection for underfill and overfill; cap inspection for presence and seating, cocked or missing closures, tamper band and seal integrity, and closure colour — the classic beverage and FMCG end-of-line checks on bottles, cans and pouches.
Presence, absence & counting
Missing components, correct assembly, item count in a pack, tablet or blister count, and completeness verification before cartoning or palletising.
Empty-container & preform inspection
Foreign objects, residue and damage in returnable bottles before filling; neck finish, thread and base defects on new containers and PET preforms.
Carton, case & pallet verification
Case count and pattern, correct case code and label, open flaps, and pallet-label reading at the end of the line so the traceability record matches what actually left.
Web & continuous-surface inspection
Line-scan cameras for film, sheet, extruded profile, coated steel, tiles and gypsum board — surface defects, holes and edge faults on material that never stops moving.
Because we design the algorithms rather than buy a fixed appliance, we can build a check for almost any visible characteristic of your product. If you can see the defect, we can usually teach the line to see it.
Where vision inspection pays in South African plants.
The check that pays back fastest is the one that stops a defect reaching a customer who charges you for it — a retailer's chargeback, a pharmaceutical recall, an OEM's rejected batch. The table is a starting point; the scoping call confirms which check on your line carries that cost.
| Industry | The checks that pay first | Typical station |
|---|---|---|
| Food & beverage — bottling, canning, dairy, bakery | Fill level, cap and tamper band, label presence and SKU, date and lot code OCV, empty-bottle inspection, foreign-object checks | Multi-camera inline station after the filler and labeller, with air-blast reject |
| FMCG & consumer packaging | Label position and artwork mix-up, seal integrity on pouches and sachets, carton print and count, case label verification | Single or dual camera per check point; end-of-line case verification |
| Pharmaceutical & nutraceutical | Blister fill and tablet count, broken or discoloured tablets, batch and expiry OCV, leaflet presence, carton serialisation code reading | Blister inspection at the sealing station; OCV at the cartoner |
| Plastics & packaging | Short-shots, flash and sink marks on moulded parts, preform defects, film and sheet surface inspection, print registration | Camera at the mould exit or take-out robot; line-scan for film |
| Building materials — tiles, board, blocks | Surface cracks, chips and colour variation on tiles and gypsum board, dimensional checks on blocks and pavers, print and grade sorting | Line-scan or multi-camera gantry over the conveyor with sort or reject |
| Metal fabrication & components | Hole and weld presence, dimensional measurement, surface defects on coated sheet, direct part mark (DPM) code reading for traceability | Fixed camera at the press or weld cell; DPM reader on the part flow |
| Agri-processing — grain, oil, feed | Bag label and weight-tag verification, seal and stitch checks, foreign-object detection on inspection belts, colour sorting | Camera at the bagging line outfeed; sorting station on the belt |
| Warehousing & sortation | Barcode and label reading on parcels and cases, dimensioning, no-read handling, pallet-label capture | Multi-side code-reading tunnel over the sorter |
Twelve situations we are asked about, and what each one actually needs.
Most enquiries start with a problem, not a technology: a chargeback, a returned pallet, a customer audit. Find the row closest to yours. The band is the cost tier from the table further down; the last two columns are what decides whether the station works — what your line has to provide, and where to read more before the call. Bands are indicative, USD, 2026.
| The situation | What to inspect | The station | Indicative band | What your line must provide | Read first |
|---|---|---|---|---|---|
| Retailer chargebacks for missing or illegible date and batch codes on bottled water, juice or dairy | Date and lot code OCV at the labeller or coder outfeed | Single smart camera with air-blast reject | USD 7,500–19,000 | Encoder pulse, PLC I/O, a stable code position, a short straight conveyor run | Food & beverage |
| Underfilled bottles reaching customers, overfill eating margin on a 2,000–6,000 bph line | Fill-level check by backlit silhouette, per container | Single camera; combined with cap and label checks on a multi-camera station | USD 7,500–19,000 alone; USD 22,000–56,000 combined | Container spacing at the check point, reject lane after the filler | Water bottling plant cost |
| Cocked caps, missing tamper bands and wrong closure colour after the capper | Cap presence, seating and tamper-band inspection, two or more views | Multi-camera inline station | USD 22,000–56,000 | Diffuse dome lighting for shiny caps, reject-verify sensor | PET, glass and can lines |
| Wrong label or SKU mix-up on a line that runs several products | Label presence, position, skew and artwork match against the recipe | Smart camera per side, or a PC-based station where artwork matching needs it | USD 7,500–19,000 per side | Recipe selection from the PLC or HMI on changeover | Food & beverage |
| Short-shots, flash and sink marks on injection-moulded caps, closures and containers | Surface and shape defect detection, deep learning where the defect varies | PC-based station at the mould take-out or on the outfeed conveyor | USD 22,000–56,000 | Consistent part presentation, labelled good and bad samples | Plastics & packaging |
| Cracked, chipped or off-colour tiles and gypsum board leaving the kiln or dryer | Line-scan surface inspection with grade sorting | High-speed line-scan gantry with sort or reject | USD 62,500+ | Encoder-tracked conveyor, strobed line light, sort mechanism | Building materials |
| Missing holes, welds or inserts leaving a press or weld cell for an OEM customer | Presence, absence and dimensional checks against the drawing | Single fixed camera, telecentric optics where tolerance demands it | USD 7,500–19,000 | Fixture or stop position, a trigger from the press cycle | Metal fabrication |
| Machined or cast parts need unit-level traceability for an automotive or mining OEM | Direct part mark (DPM) code reading and grading | Fixed-mount DPM reader with result to MES | USD 7,500–19,000 | Marking that meets the customer's code standard, MES or database endpoint | Press and CNC lines |
| Blister packs shipped with missing, broken or discoloured tablets | Blister fill and tablet count at the sealing station, expiry OCV at the cartoner | Multi-camera PC-based station | USD 22,000–56,000 | Access at the sealer, batch data from the line controller | Line upgrade & digitalisation |
| Wrong case count or unreadable pallet label found by the distributor, not by you | Case pattern and count, case-label reading, pallet-label capture at the end of the line | Multi-side code-reading tunnel | USD 22,000–56,000 | Label placement rules, WMS or ERP endpoint for the record | OEE on an old line |
| Returnable glass bottles with residue or foreign objects entering the filler | Empty-bottle inspection: base, sidewall, finish and residual liquid | High-speed multi-camera station before the filler | USD 62,500+ | Single-file bottle flow, reject before the rinser or filler | Food & beverage |
| Feed, flour or grain bags leaving with the wrong weight tag or an open stitch | Bag label and weight-tag verification, stitch and seal check | Single camera at the bagging line outfeed | USD 7,500–19,000 | Bag orientation at the check point, reject pusher | Agri-processing |
If your situation is not in the table, send the samples anyway: the imaging trial answers in a day whether a camera can see the defect, and the row it belongs to follows from that.
Vision, or something else? Matching the check to the tool.
Not every quality problem is a camera problem. The table is how we sort a request before quoting, and why a vision assessment sometimes ends with a recommendation for a cheaper instrument.
| Machine vision station | Checkweigher | Metal detector / X-ray | Sampling by an inspector | |
|---|---|---|---|---|
| Finds | Anything visible: defects, codes, labels, fill line, caps, dimensions, count, colour | Under- and over-weight packs, missing items by mass | Metal fragments; dense foreign bodies and gross fill faults with X-ray | Attributes that need judgement — taste, feel, complex assembly |
| Coverage | 100% of production, every result logged | 100%, weight only | 100%, contaminants only | A sample; results depend on the inspector and the hour |
| Misses | Hidden internal faults, weight, contaminants inside opaque packs | Visible defects, wrong label, code errors | Cosmetic and print defects, labels, codes | Everything between samples |
| Best used | Where the defect is visible and the cost of shipping it is high | Where fill by mass is the regulated or costly variable | Where contamination is the risk, usually alongside vision | Low volume, high variety, or as the audit on top of automated checks |
| Typical pairing | On a packaging line the three instruments sit in sequence — vision at the labeller and capper, checkweigher after the filler, metal detector or X-ray before case packing — and all three hand results to the same PLC and reject lane | |||
Built on HIKROBOT VisionMaster.
Our inspection systems are built on VisionMaster, the machine-vision software platform developed by HIKROBOT (Hangzhou Hikrobot Co., Ltd., the machine-vision and mobile-robot subsidiary of Hikvision), paired with HIKROBOT industrial cameras, lighting and vision controllers. VisionMaster gives us a deep algorithm library for positioning, measurement, defect detection and code and character recognition, three ways to build — a graphical drag-and-drop environment, an SDK for custom development, and an operator-design mode — and a separate deep-learning package for segmentation, object detection, classification, OCR and anomaly detection.
What matters to you is not the brand but what it lets us do: develop a custom inspection recipe quickly, run it on a smart camera or a vision controller depending on the job, talk to your PLC over the protocols your line already uses, and hand you a station your own team can operate. The same logic applies as with our IoT and OEE work: we choose a strong platform, then add the integration, algorithm design, and on-site engineering that makes it actually solve your problem. See our technology partners.
What you own: the configured inspection recipes, the result data, and full documentation. We build it so your team can run it — and so the results export cleanly into your OEE and quality systems.
The HIKROBOT machine-vision hardware range CISH builds on.
| Layer | What HIKROBOT provides | What CISH does with it |
|---|---|---|
| Software | VisionMaster algorithm platform (location, measurement, identification, defect detection); deep-learning package; SC_VisionMaster on smart cameras; a separate 3D platform and RobotPilot for vision-guided robots | Designs the inspection logic, validates it on your samples, builds the operator interface and recipe management |
| Cameras | Area-scan (GigE and USB3), line-scan, smart cameras (SC1000–SC6000 series), fixed and handheld code readers, 3D laser profile sensors, stereo 3D cameras | Chooses sensor, resolution and interface for the field of view, speed and defect size |
| Optics & lighting | FA and M12 lenses, light sources and light controllers | Designs the imaging so the defect shows and the noise doesn't — the part that decides whether a system works |
| Compute & I/O | VC-series vision controllers with opto-isolated I/O for PLCs and reject actuators; GigE Vision, USB3 Vision and GenICam-compliant cameras | Integrates the station with your PLC, encoder and reject mechanism, and logs results |
| Communication | TCP/IP, Modbus, serial, UDP and EtherNet/IP on VisionMaster; PROFINET, EtherNet/IP and Modbus on smart code readers (vendor documentation) | Maps results and reject signals to the protocol your PLC speaks; feeds OEE and MES |
Platform capabilities are as described in HIKROBOT's published product documentation (2025–2026); see primary sources below. GigE Vision, USB3 Vision and GenICam are open camera-interface standards, so the station is not locked to one camera brand.
Smart camera, PC-based station, or vision sensor?
The same check can be built three ways, and the price and flexibility differ by an order of magnitude. Choosing wrong in either direction costs money: a vision sensor that cannot cope with product variation, or a PC-based system for a job a smart camera does for a fraction of the cost.
| Vision sensor | Smart camera | PC-based vision system | |
|---|---|---|---|
| What it is | One camera, fixed function, taught by example | Camera with the processor and VisionMaster tools on board (SC_VisionMaster) | One or more cameras connected to a vision controller running the full VisionMaster platform |
| Best for | Presence/absence, cap on, label on — one simple pass/fail | One or two checks at a single point: OCR/OCV, code reading, position, a defined defect | Multi-camera, multi-check stations; deep learning; measurement; high line speeds; several inspection points sharing one controller |
| Integration | Digital I/O to PLC | Digital I/O plus industrial Ethernet to PLC | Full PLC, MES and database integration; encoder-tracked reject |
| Indicative cost | Lowest | USD 7,500–19,000 as a single-camera station | USD 22,000–56,000 for a multi-camera inline station; USD 62,500+ for high-speed or multi-station |
| Watch out for | No margin for product variation; no data | Processing limits at high speed or with deep learning | Over-specifying a simple check |
When AI defect detection is worth it — and when it is not.
Rule-based tools — edge, blob, pattern matching, measurement — are fast, deterministic and easy to explain to a quality auditor. They are the right choice for anything you can describe: a dimension, a code, a label position, a fill line. Deep learning earns its place when the defect cannot be described, only recognised: scratches and dents that vary in every instance, contamination on a natural product, cosmetic defects on textured surfaces, characters printed on curved or deformed packaging.
VisionMaster's deep-learning package covers classification, object detection, segmentation and OCR, plus anomaly detection that trains on good samples only — useful when you have thousands of good parts and a handful of bad ones. The practical questions are how many labelled samples you can supply, how stable the imaging is, and how you will handle the false rejects while the model matures. We start with rules where rules work, add a model only for the checks that need it, and keep a quarantine lane rather than scrapping on a model's say-so during the first weeks.
Sample discipline
Good and defective samples from real production, across shifts, suppliers and seasons — not a lab set. Labelled with the same criteria your quality team uses.
Explainability
Defect heat-maps and stored images for every reject, so an auditor or a supplier can see why a part failed.
Retraining path
New SKU, new supplier, new defect: your team adds samples and retrains without calling us for every change.
Lighting and optics decide whether a vision system works.
Most failed vision projects failed in the imaging, not in the software. A defect the camera cannot see cannot be found by any algorithm, and a defect that only shows under the morning sun through the roof lights will produce rejects every afternoon. Imaging design is the part of the job we spend the most time on.
Lighting technique
Backlight for silhouettes and fill level; diffuse dome for shiny caps and foil; dark-field for scratches; coaxial for flat, reflective surfaces; colour and IR to separate print from substrate; strobed line light for web inspection.
Lens and working distance
Resolution sized to the smallest defect, telecentric lenses for measurement without perspective error, line-scan optics matched to conveyor speed and encoder pulses.
Plant conditions
IP-rated enclosures and air purge for dust, washdown-rated housings for food lines, shrouds against ambient light, vibration-isolated mounts on presses and gantries.
PLC, reject mechanism and traceability — the station has to fit the line.
An inspection station is only useful if the reject happens at the right product, every time, and the result lands where your quality system can use it. That is an integration job, not a camera job.
- PLC hand-off. Pass/fail and measurement results go to your line PLC over digital I/O or industrial Ethernet — TCP/IP, Modbus, EtherNet/IP, and PROFINET where the hardware supports it — with a heartbeat so a stalled station stops the line rather than passing everything.
- Reject tracking. Encoder-tracked shift registers so the air-blast, pusher or diverter fires on the failed product at line speed, with a reject-verify sensor to confirm it left the line.
- Fail-safe logic. No result means reject or stop, never pass. Camera or lighting faults raise an alarm.
- Recipes per SKU. Operators select the product on the HMI or the PLC sends the recipe; changeover takes seconds, not a service call.
- Results and images. Every inspection logged with a timestamp, batch and result; reject images stored; counts and reject reasons exported to OEE, MES or a database — see OEE measurement on an old line.
Retrofit on an existing line: most stations are added to a running line as a self-contained module — cameras, lighting, controller and reject — interfaced to the existing PLC. You do not need to replace your line controls, and an old PLC is rarely the obstacle; see when to upgrade your PLC. Vision inspection is a natural first step in a line upgrade and digitalisation programme.
False rejects and missed defects: the two numbers we sign off on.
Every vision system makes two kinds of mistake. A false reject throws away good product — the cost is scrap, rework and operators who learn to override the station. A missed defect (escape) lets bad product through — the cost is the customer, the recall, the chargeback. The two trade off against each other through the inspection threshold, so a system is only "working" when both are inside limits you agreed before commissioning.
Agreed targets
False-reject rate and escape rate are written into the scope with the line speed, before we choose a camera.
Golden and defect sets
A retained set of known-good and known-bad samples is run at commissioning and after every change, so the numbers are measured, not assumed.
Tuned on real production
Thresholds are set on your product, your shifts and your suppliers' variation, and reviewed in the first weeks of running.
Vision systems on real production and sortation lines.
From defect sample to a tuned, running station.
Define the defect
We work from real product samples — good and bad — to define exactly what must be caught, the acceptable false-reject rate, and the line speed it has to run at.
Design the imaging
Camera, lens, and — most importantly — lighting are chosen for your product and defect. Good lighting is most of a reliable vision system; this is where most cheap systems fail.
Build the algorithm
We develop the inspection logic on VisionMaster — rule-based tools first, deep learning where the defect needs it — then validate it against a representative set of real good and defective samples.
Integrate & reject
The station is interfaced to your line PLC, with a reject mechanism (air-blast, pusher, diverter) and result logging into your OEE / quality data.
Tune on real production
We commission on your actual product and shift conditions, driving the false-reject and missed-defect rates down to agreed targets before sign-off.
Train & hand over
Your operators learn to run recipes, add product variants, and read the results. Documentation and recipe ownership stay with you.
What an inspection station typically costs.
| System | Indicative cost (USD) | Typical use |
|---|---|---|
| Single-camera station | USD 7,500–19,000 | One check — e.g. code verification, label presence, a single defect type |
| Multi-camera inline station | USD 22,000–56,000 | Several checks at once, custom algorithms, full line integration and reject |
| High-speed / multi-station system | USD 62,500+ | High line speeds, 360° inspection, multiple inspection points, complex defects |
Cost is driven by camera count, line speed, lighting complexity, and algorithm difficulty — not by a fixed price list. We scope to your actual product and defect. A free assessment usually establishes the right band quickly.
What drives the price up
- Line speed and part rate — more frames per second, faster processing, line-scan instead of area-scan.
- Number of views — 360° inspection of a round container needs several cameras or mirrors.
- Defect subtlety — small, low-contrast or variable defects need higher resolution, more lighting engineering, or deep learning.
- Measurement accuracy — telecentric optics, calibration and 3D profiling cost more than a presence check.
- Environment — washdown, dust, heat and vibration add enclosures and mounts.
How to think about payback
Add up what the defect costs you now: scrap and rework, customer returns and chargebacks, the inspectors on the line, and the occasional recall or lost contract. Set it against the station's cost band above. Where a check protects a customer relationship or a regulated product, the payback is usually measured in months; where it replaces a sampling inspector on a slow line, it may not pay at all — and we will say so in the assessment. The full arithmetic is in machine vision inspection: cost and where it pays.
A machine vision integrator in Johannesburg, working across South Africa and Africa.
Most vision hardware sold in South Africa comes through distributors who supply components and expect you, or a systems integrator, to make them work on the line. CISH is the integrator: we take responsibility for the check, not the camera. Our engineers are based in South Africa, so the site survey, the imaging trials on your product, the commissioning and the follow-up tuning happen on your floor, and the station is supported by the same people who built it. For plants elsewhere in Africa, the same team designs and commissions through our sister site, cish.africa.
Site survey and imaging trial
We image your real product and defect samples before quoting, so the proposal is based on what the camera can actually see.
Commissioning on your line
Installation, PLC integration, reject tuning and acceptance against the agreed false-reject and escape targets, on your shifts.
Support and changes
New SKUs, new defects, camera or lighting replacements, and recipe changes handled by the team that built the station — with your operators trained to do the routine ones themselves.
Questions buyers ask about vision inspection.
The questions plant managers, quality managers and engineers ask us most often, answered plainly.
On the quote, buying direct is usually cheaper; for the project as a whole, not necessarily. The quote assumes you already have six capabilities — specification, supplier screening, negotiation, shipping and duties, commissioning, maintenance — and can carry the risk of the project failing. What those cost you, in time and in tuition, is the real comparison: Is buying through CISH more expensive?
Vision inspection in context.
Machine vision inspection: cost & where it pays
What it costs, AOI vs manual inspection, and the checks with the fastest payback.
Cost to digitalise a production line
Where vision inspection fits in a wider digitalisation budget.
Line Upgrade & Digitalisation
Vision inspection as part of a broader line-upgrade and OEE programme.
Got a defect you keep shipping?
Send us a few good and bad samples, the line speed and where the check would sit. We will image them, tell you whether a camera can see the defect, and come back with the right architecture and cost band — including when a simpler sensor is all you need.
Primary sources
- HIKROBOT — VisionMaster machine vision software platform (algorithm tools, development modes, deep-learning modules)
- HIKROBOT — Download centre: VisionMaster V4.4.0, deep-learning package, SC_VisionMaster, 3D platform and RobotPilot releases
- HIKROBOT — Machine Vision Smart Product Catalog 2025Q4 (communication protocols, GigE Vision / USB3 Vision compatibility, smart camera and reader families)
- HIKROBOT — VC3000 series vision controller datasheet (opto-isolated I/O for PLC and actuator wiring)
- HIKROBOT — GigE line-scan camera datasheet (GigE Vision V2.0 and GenICam compliance)
- HIKROBOT — About Hangzhou Hikrobot Co., Ltd. · Hikvision — 2025 Annual Report (Hikrobot as a subsidiary; VM and RobotPilot software)
- EMVA — GenICam standard (transport-agnostic camera interface across GigE Vision, USB3 Vision, CoaXPress, Camera Link)
- gov.za — Media statement: new regulations on labelling and advertising of foodstuffs (R.146 of 1 March 2010) become law