TL;DR
Machine vision AI uses industrial cameras, controlled lighting and trained AI models to inspect, read, measure or guide parts on a production line. A sensor triggers the camera, and a model running at the edge classifies or locates what it sees. The station then sends a pass, fail or review decision to the PLC, which rejects the part and logs the result. Rule-based vision still suits stable, measurable features, while deep learning handles natural variation, and a single station often uses both. Success depends less on the model than on lighting, real defect images, the line’s timing budget and links to MES and quality systems. Prove it on one station with a measured baseline, then check escapes and false rejects per defect class before you scale.
Key Takeaways AI-driven machine vision pairs industrial cameras, lighting and optics with deep learning models that inspect, read, measure or guide parts on a production line. Rule-based vision still wins for stable, measurable features, while deep learning earns its place where appearance varies, and stations often combine both. Lighting and optics decide what the model can see, and an accuracy problem often turns out to be an imaging problem. Inference runs at the edge because the decision has to reach the reject actuator inside the line’s cycle time. Escapes and false rejects need to be measured per defect class, because a single accuracy number hides the costly failures. The safest start is one bounded station with a baseline, an owner and a go/no-go gate before scaling. Watch on YouTube
Optimizing Production & Supply Chain Through AI Implementation | Kanerika
Kanerika walks through how a producer put AI into production planning and supply chain decisions, the plant-floor data work that any inspection program also depends on.
The Same Inspection Line, Before and After AI
On the old version of the line, a rule-based camera watches brushed-steel housings roll past. Then a new coil of steel arrives with a slightly different grain. Suddenly the threshold that caught scratches yesterday rejects clean parts by the hundred. So an engineer loosens it, the false rejects stop, and a faint scratch rides through to a customer.
On the upgraded line, the camera and the light are the same. This time, however, a model trained on images of real scratches and real grain makes the call. As a result, a change in grain no longer looks like damage. That shift is what AI brings to machine vision on a factory floor.
Keeping that new judgement reliable for months is the hard part. It depends on lighting, data, timing and validation far more than on the model itself, and this guide walks through each one.
What Is Machine Vision AI?
Machine vision AI is industrial machine vision whose decisions come from trained models rather than only from hand-set rules. TechTarget defines machine vision as a computer’s ability to see through cameras, analog-to-digital conversion and signal processing. The resulting data then goes to a computer or robot controller. The same definition notes that machine vision is often combined with AI, machine learning and deep learning.
So is machine vision AI by definition? Not always. A classic system that measures an edge with a caliper tool involves no learning at all. It only becomes AI when a trained model decides what the camera is looking at.
How AI Changes Traditional Machine Vision
Traditional machine vision runs on explicit rules. An engineer sets a brightness threshold, finds edges, matches a template or measures a gap. The tool then returns a number to compare against a tolerance. That approach works well when parts look the same every time.
Trouble starts when appearance varies naturally. Wood grain, cast surfaces, textured plastics and printed labels change from part to part, so a fixed rule either misses defects or rejects good parts. A deep learning model instead learns what acceptable variation looks like from labeled examples, which is why it copes where rules break.
Talk to Kanerika
Rules, Deep Learning or Both for Your Inspection?
Kanerika reviews your parts, defect types and line timing, then tells you where a rule-based tool is enough and where a trained model is worth the data work.
Talk to an Expert → Machine Vision vs Computer Vision: Where AI Fits
Computer vision is the broad field of teaching software to interpret images, from phone cameras to medical scans. Machine vision is its industrial branch, built for fixed stations, controlled light and hard timing. Machine vision AI is that branch with learned models doing the judging, and it shares much of its modelling toolkit with AI image recognition .
Table 1: Computer vision, rule-based machine vision and AI-driven inspection compared
Aspect Computer vision Machine vision (rule-based) AI-driven machine vision Environment Any images, any setting Fixed station, controlled light Fixed station, controlled light Main goal Understand image content Measure and check against set rules Judge variable parts the way a trained inspector would Typical output Labels, captions, boxes Measurements and pass/fail Defect class, location and pass/fail/review Timing Often seconds or batch Milliseconds, inside cycle time Milliseconds, inside cycle time Example Tagging product photos Gauging a hole diameter Finding scratches on textured metal
How Machine Vision AI Works on a Production Line
Every inspection follows the same loop, and the whole loop has to finish before the part reaches the reject point. A typical station runs five steps.
Trigger. A photo-eye or conveyor encoder tells the camera a part is in position, so every image is taken at the same spot.Capture. The camera exposes under controlled light, and the image is cropped, aligned and normalised so the model gets consistent input.Infer. A trained model analyses the image on an edge device next to the line and returns a class, a location or an anomaly score.Decide. Rules turn that output into pass, fail or manual review, using thresholds that quality engineers own.Reject or log. The PLC fires the reject actuator if needed, and the station records the image, the result and the model version.Inside the infer step, the model usually does one of four jobs. Classification labels the whole image as good or as a defect type. Object detection draws boxes around parts or flaws, while segmentation outlines the exact defect pixels so their size can be measured.
The fourth job, anomaly detection , learns only from good parts and flags anything unusual. As a result, it suits defects too rare to collect in quantity. Convolutional neural networks remain a common backbone for all four tasks.
Running these models on edge AI hardware is what keeps the loop inside the line’s cycle time. In the end, the station produces a disposition the line can act on, not a prediction for someone to read later.
Machine Vision AI Use Cases in Manufacturing
The same loop shows up in six common inspection jobs. Each one has a place where AI clearly helps and a place where older tools still do better.
Surface Defect Detection
Scratches, dents, porosity, stains and coating flaws are the classic AI case. Their shape and contrast vary too much for a fixed threshold, so a trained model separates real damage from harmless texture where a rule struggles.
Assembly Verification and Presence Checks
Is the clip seated, is the connector fully home, are all six screws in place? Simple presence checks, for example, often work fine with rules. AI helps when parts arrive at odd angles, cables move, or the background changes between product variants.
Dimensional Measurement
Gauging a hole diameter or a gap against a tolerance is still a job for calibrated, rule-based tools. AI can find the feature on a messy part first. Even so, the measurement itself should come from the calibrated tool, because a model’s estimate is not traceable to a standard.
OCR, Code Reading and Label Inspection
Clean barcodes and 2D codes need a decoder, not a neural network. Deep learning OCR earns its place on embossed, laser-etched or curved text, and on labels where print quality drifts. Label checks also confirm that the right artwork, lot code and expiry date sit on the right product.
Robot Guidance and Pick-and-Place
Here the camera locates a part so a robot can grip it, often from a bin of randomly piled pieces. Vision is only one input in that setup, and the wider story of machines that perceive and then act belongs to physical AI .
Packaging and Seal Inspection
Seal integrity, fill level, cap position and carton print are all checked at high speed on packaging lines. Pharma packaging adds serial number and print checks, so the image record matters as much as the decision.
The same ideas travel across industries, although the constraints change. For instance, chipmakers apply them to wafer and die inspection, as our guide to AI in semiconductor manufacturing explains. Outdoor sites are a different problem, and AI in construction deals with cameras facing weather and shifting light instead of a fixed station.
Retail shelves follow their own pattern, which our post on computer vision in retail covers. For the plant-wide picture beyond inspection, see our guide to AI in manufacturing .
Kanerika Service
AI for Manufacturing
Kanerika brings AI, data engineering and analytics to plant operations, from predictive maintenance to quality control models built on your operations data.
Explore AI for Manufacturing What AI Inspection Changes on the Line
The gains come from mechanisms you can name and measure, not from a headline percentage. Five stand out.
Consistent judgement. The model applies the same standard at 3 a.m. as at noon, on every shift.Full inspection instead of sampling. Every part is checked at line speed, so a bad batch surfaces early.Tolerance for natural variation. Changes in grain, texture and finish stop triggering false rejects.Traceability. Each image record ties a defect to a lot, a shift and a station, which speeds up root-cause work.Better use of inspectors. People adjudicate borderline cases and grow the defect library instead of staring at parts all day.How much each gain is worth depends on your own baseline. For that reason, measure escapes, scrap and inspection hours before the project starts, then use the cost section below to build the case.
Checklist
Enterprise AI Checklist
Review strategy, data and platform readiness, ownership, governance and deployment for an AI use case before it scales.
Get the Checklist → Rule-Based vs Deep Learning vs Hybrid: Choosing the Method
Choosing the method is the first real design decision for any station. The answer is rarely all one way, and Table 2 shows why.
Table 2: Rule-based, deep learning and hybrid inspection compared
Factor Rule-based Deep learning Hybrid Best for Stable, measurable features Variable appearance, cosmetic defects Mixed jobs on one part Data needed A few setup images Labeled good and defect images Labeled images for the AI step only Explainability Every decision traceable Needs heatmaps and test evidence Final accept logic stays explicit Changeover effort Retune rules per variant Collect images and retrain Usually only the AI step changes Compute Modest Edge accelerator or GPU Edge accelerator or GPU Typical failure False rejects when appearance shifts Silent drift on unseen conditions Integration gaps between steps
When Rule-Based Tools Are Enough
Rules win when the feature is stable, measurable and well lit. For example, gauging, alignment, presence checks on rigid parts and code reading all fall here. These tools need no training images, explain every decision and run on modest hardware.
When Deep Learning Earns Its Place
Deep learning pays off when good parts look different from each other and defects vary in shape. Even so, it needs labeled images, more compute and a plan for retraining. In exchange, it handles texture, reflections and part-to-part variation that would otherwise take endless rule tuning.
Why Stations Often End Up Hybrid
In practice, a rule-based tool often locates and aligns the part first. Next, a deep learning model classifies any defect, and rule-based accept logic applies the tolerance. Vendors build for this pattern too, and MVTec’s deep learning overview describes AI combined with rule-based machine vision in production inspection, including anomaly detection.
The takeaway for a project team is to drop the idea that AI always beats rules. Use each method where it is strongest, and keep the final accept logic explicit so quality engineers can read it.
Camera, Lighting and Optics: Where Inspection Accuracy Is Won
A model can only judge what the image shows. If a scratch is invisible in the image, no amount of training will find it, so imaging design comes before model design.
Resolution From the Smallest Defect
Start from the smallest defect you must catch and work backwards. A common rule of thumb, not a standard, puts three or four pixels across that defect. Take a hypothetical 100 mm field of view and a 0.2 mm defect.
That field is 500 defect-widths across. At three to four pixels each, you need roughly 1,500 to 2,000 pixels across the sensor, before any margin for part position.
Lighting Techniques That Make Defects Visible
Lighting decides contrast, and contrast decides whether a defect exists in the data at all. Table 3 lists four common techniques and what each one reveals.
Table 3: Four lighting techniques for machine vision inspection
Technique What it reveals Typical parts Backlight Silhouettes, edges and holes Stamped parts, gaskets, fill level in clear containers Bright-field Flat surface detail and print Labels, flat machined faces Dark-field Scratches, dents and raised marks Polished metal, glass, engraved parts Diffuse dome Detail under glare on curved, shiny parts Cans, bearings, foil packs
Polarizing filters can also cut glare on wet or shiny surfaces. Once the setup works, lock it down, because a replaced lamp or a nudged light changes every image the model sees.
Motion, Shutters and Triggering
Because moving parts blur, they need a global shutter, which exposes the whole sensor at once, or a short strobe that freezes motion. Encoder-based triggering then keeps the part in the same position in every frame, so images stay comparable.
When 2D Is Not Enough
Height, flatness and solder volume need 3D sensing, such as laser profiling or structured light. Continuous material like film, foil or textiles usually calls for line-scan cameras instead of area cameras.
When you compare cameras, use their EMVA 1288 data, the standard for objective characterization of industrial cameras and image sensors. Still, a good camera is not the same as a good inspection, which is why the setup above matters more.
Training Data That Reflects Real Factory Variation
The model learns the factory you show it. If the training set comes from one good week on one shift, the model will be surprised by everything else.
Define Defects Before Anyone Labels
Quality engineering should own a written defect taxonomy with names, example images and severity levels. Borderline samples then go to an adjudication group, so two labelers do not call the same scratch differently. Good data labeling tools help, but the definitions come first.
Plants with several sites also need shared defect codes, which makes this a data standardization task as much as a vision one.
Capture Variation Across Lots, Shifts and Changeovers
First, collect images across material lots, shifts, seasons and product changeovers. Then split training and test sets by lot or by time, never by random frames. Consecutive frames are near-duplicates, so a random split leaks test images into training and inflates the score.
Rare Defects and Anomaly Detection
Some defects appear only a few times a year. Anomaly detection trained on good parts can flag them, while a targeted collection effort builds real examples over time. Synthetic images can help training, but validation should always use real defects.
Traceable Image Records
Store every image with its lot, station settings, model version and final disposition. Those records feed AI data quality checks, retraining and audits.
They also give you the evidence to manage model risk under a frame such as the NIST AI Risk Management Framework , released on January 26, 2023.
Datasheet
Manufacturing AI Readiness: Critical Gaps and a Path Forward
A datasheet on the gaps in data trust, legacy platforms, manual processes and workforce readiness that hold manufacturing AI back, and practical ways to close them.
Get the Datasheet → Edge Deployment: Meeting the Line’s Cycle Time
A cloud round trip is too slow and too fragile for a reject decision. Inference therefore runs beside the line, on a smart camera, an industrial PC or a GPU edge box.
Budget the Full Image-to-Reject Time
The model is only one slice of the time available. Take a hypothetical line where 120 ms separate the camera from the reject gate. Exposure and transfer might take 15 ms, preprocessing 10 ms, inference 25 ms and the decision 5 ms.
The PLC message and actuator travel could then use another 40 ms, which leaves 25 ms of margin. Measure each slice on the real hardware, because a model that is fast on a workstation GPU can be slow on an embedded chip.
Table 4: What runs at the edge and what runs in the cloud
Responsibility Edge, at the line Cloud or data center Inference and reject decision Always Never in the control path Model training Rarely On pooled images from many stations Model versions Runs the approved version Registry, approvals and rollback Analytics Live station view Cross-line and cross-site trends
Smart Cameras, Industrial PCs and GPU Edge Boxes
Smart cameras now run AI on board. Cognex launched its In-Sight 3900 on May 5, 2026, built on Qualcomm Dragonwing platforms, with inspections at the edge and no external PC. Industrial PCs and GPU boxes suit heavier models or several cameras at once.
Under the hood, interfaces such as GigE Vision and USB3 Vision sit beneath the GenICam programming interface, which keeps software portable across camera hardware. Our guide to edge AI for real-time analytics covers the wider edge stack.
Offline Behavior and Fail-Safe Modes
Above all, the station must keep inspecting when the network drops. When the model is unsure or the camera fails, the part should go to manual review, never pass by default.
Vision PCs and cameras are also operational technology assets. Secure them with NIST SP 800-82 Rev. 3 , the OT security guide published in September 2023, while noting that Revision 4 exists only as a public draft.
Inspection Is Not a Safety Function
A quality model decides whether a part is good. It should never decide whether a person is safe near a machine. Safety-related parts of control systems are designed and validated to standards such as ISO 13849-1:2023 . A quality-inspection model is not built that way, so keep vision inspection outside that safety chain.
Connecting Inspection Results to PLC, MES and Quality Systems
An inspection result only creates value once the right system acts on it. ISA-95 , also known as IEC 62264, is the standard set for integrating enterprise systems with manufacturing control. Its level model shows where each decision belongs.
Keep Control Decisions at the Line
The reject itself stays at the control level, between the vision station and the PLC. Inspection events then travel up to MES and quality systems, while analytics sit above them. As a result, an outage higher up never stops a reject on the line.
In practice, one inspection event climbs four tiers. It starts as a reject at the PLC, becomes a record in MES and QMS, joins the defect history in a lakehouse and finally shows up on a dashboard.
An Inspection Event Schema
Agree on one event format early, so every station speaks the same language. Keep the field names identical across stations, or the analytics layer ends up reconciling formats instead of finding defects. Here is a minimal example with made-up values.
{
"part_id": "HX-204-118734",
"station": "line3-cam2",
"timestamp": "2026-11-06T09:14:22.418Z",
"defect_class": "scratch",
"confidence": 0.91,
"disposition": "fail",
"image_uri": "s3://inspection/line3/cam2/118734.png",
"model_version": "surface-v14"
}Each field earns its place. The model version and image link, for example, let an engineer replay any decision months later. Similarly, the confidence value shows which calls sat close to the threshold, so borderline parts can be pulled for review. Together, these fields turn each reject into evidence instead of a mystery. A real deployment would add a lot number and a shift code, since root-cause reviews group rejects by both.
From Detection to Quality Workflow and Analytics
A fail event can open a nonconformance record in the QMS, while a run of the same defect alerts the line lead. Over time, events land in a data lakehouse for manufacturing , where dashboards show yield by line, a defect Pareto and drift by station. Tagging each event with its lot and shift also lets engineers trace a defect spike back to a material batch or a changeover. Likewise, a false-reject rate that creeps up on one station can point to a worn light or a shifted lens before scrap climbs.
That is also where manufacturing analytics connects inspection to scrap, rework and supplier quality, so the cost of a defect sits next to the count. Quality teams can then compare stations side by side, and supplier issues show up as patterns instead of isolated complaints. Because of that link, a vision station becomes a data source for the whole plant.
Watch on YouTube
How Business Intelligence is Transforming Modern Manufacturing Plants?
A look at how plants turn line and quality data into dashboards people act on, which is where inspection events end up once they leave the station.
Validating an AI Inspection Model Before It Controls Release
A single accuracy figure hides the failures that cost money. Instead, measure two errors separately, namely escapes (bad parts passed) and false rejects (good parts scrapped or reworked).
Escapes and False Rejects, Measured Separately
Consider a hypothetical test set of 10,000 parts that contains 50 real defects. A model that passes everything scores 99.5% accuracy, yet all 50 defects escape. Even a model that catches half of them still ships 25 bad parts while looking excellent on paper.
For that reason, set thresholds with quality leadership by defect severity. A cosmetic mark and a cracked seal deserve very different trade-offs.
Test by Defect Class and Condition
Next, report results per defect class, per product variant and per condition, such as each shift and material lot. Also use held-out lots the model has never seen, and keep borderline samples in a separate set that people adjudicate.
Monitor Drift and Revalidate After Changes
Models drift when the world in front of the camera changes. A lamp replacement, a bumped camera, a new supplier, a new SKU or a retrain should each trigger revalidation. Meanwhile, keep a rollback to the last approved model ready, and use an AI governance process to name who signs off.
Costs, ROI and a Go/No-Go Checklist
Hardware is the visible part of the bill, but it is only one line on it. A realistic budget counts every step from first image to signed-off rollout.
What a Station Really Costs
The full list covers cameras, optics, lighting, fixtures and edge compute. It also includes labeling time, integration, validation, software licences, maintenance and the work every product changeover brings.
Build the Business Case From Your Baseline
Base the savings on your own numbers, not on a vendor’s. Start with the cost of each escape, then add scrap and rework from false rejects and today’s inspector hours. Keep cash savings separate from redeployed capacity, because finance teams treat them differently.
Go/No-Go Before Scaling
Before a pilot station becomes a rollout, check six things.
The defect is clearly visible in the image under the chosen lighting. A named owner in quality engineering signs off on definitions and thresholds. Training and test data include real defects across lots and shifts. A measured baseline exists for escapes, false rejects and inspection hours. The full image-to-reject time fits inside the cycle time on real hardware. Fail-safe behavior and a revalidation owner are in place. If any item is missing, fix it on the pilot first. After all, scaling copies every gap onto every line.
AI Assessment
Is Your Plant Ready for Machine Vision AI?
Kanerika’s AI Maturity Assessment scores your AI/ML foundations, generative AI capabilities and AI agent readiness, so you know which gaps to close before a vision pilot scales.
Start Your AI Assessment → AI Vision Inspection Trends to Watch in 2026
Four shifts stand out this year, and each one changes how stations are built or managed.
Embedded AI on smart cameras. Cognex’s In-Sight 3900, launched in May 2026, is a recent example of AI inspection running on the camera itself, with no external PC.Multi-site model management. Cognex announced general availability of OneVision in June 2026, and Industrial Technology reports that more than 100 customers have used it since its June 2025 beta launch.AI agents that help build vision apps. MVTec announced its AI Vision Solver for VISION 2026 in Stuttgart and describes it as a forward-looking prototype, not a shipping product.Models that describe defects in words. Vision-language models can describe an image in plain text, which makes them a candidate for drafting defect notes.Market sizing, meanwhile, depends on scope. Global Market Insights estimates the industrial machine vision market at USD 6.8 billion in 2025. By contrast, ResearchAndMarkets puts the broader machine vision systems market at $15.15 billion for the same year.
How Kanerika Helps Manufacturers Put Machine Vision AI to Work
Kanerika acts as the AI, data and integration partner on a vision program, alongside the camera vendor and the controls integrator. It does not make cameras or certify machine safety. Instead, Kanerika builds the models, pipelines and system links that turn one working station into something the whole plant can rely on.
Kanerika’s Approach: Assess, Design, Build, Integrate, Govern
Assess. Check the station for a visible defect, a measured baseline and imaging feasibility.Design. Set the defect taxonomy, labeling plan and lot-based test sets with quality engineering.Build. Train a hybrid model that keeps rule-based steps where they are stronger.Integrate. Connect edge inference to the PLC, MES and QMS, then land events in a lakehouse.Govern. Monitor drift, version models and revalidate after changes through AI governance services .Kanerika’s closest proof point is visual authentication rather than a factory line. For a global luxury goods retailer, Kanerika trained FLIP to perform computer-vision-based product authentication using image recognition. It also embedded blockchain-based provenance tracking and deployed Karl for real-time verification during returns and resale.
According to the AI vision case study , the work delivered 95% accuracy in counterfeit detection, 68% faster product verification and 100% product traceability. An inspection station needs the same building blocks: a trained vision model, verification inside the live workflow and a traceable record for every item.
On the plant side, Kanerika’s AI for manufacturing practice already builds quality control models from historical operations data. Our AI services team would begin a vision program with one station. Along the way, the team watches for the usual pitfalls, such as unlogged lighting changes, random data splits, a single accuracy KPI and missing fail-safes.
Case Study
95% Accuracy in Counterfeit Detection with AI Vision
A global luxury goods retailer deployed AI vision and blockchain for authentication, reaching 95% counterfeit detection accuracy and 68% faster verification.
Read the Case Study → Wrapping Up
Machine vision AI is industrial inspection with trained models making the judgement calls that fixed rules get wrong. On the line, it runs a tight loop of trigger, capture, inference, decision and reject inside the cycle time. Holding up in production, however, depends on the work around the model.
That work means lighting that shows the defect, data split by lot, an edge time budget, events wired into MES and quality systems, and validation per defect class. So start with one station, measure it against a baseline, and scale only when it passes the gate.
Frequently Asked Questions
Is machine vision AI?
Not by definition. Machine vision is the industrial use of cameras and image processing to inspect, measure or guide parts. Classical systems run on fixed rules such as thresholds and edge tools. It becomes AI when a trained model, usually a deep learning network, makes the decision about what the camera sees.
What is machine vision in AI?
In AI, machine vision refers to applying trained models to images captured by industrial cameras on a production line. The models classify parts, locate defects, read text or flag anomalies. The result goes to a PLC or quality system as a pass, fail or review decision, usually within the line’s cycle time.
What is the difference between machine vision and computer vision?
Computer vision is the broad field of teaching software to interpret images in any setting, from phones to medical scans. Machine vision is its industrial branch, built for fixed stations, controlled lighting and strict timing. Machine vision systems produce decisions a production line acts on, such as rejecting a defective part.
How does AI machine vision work in manufacturing?
A sensor or encoder triggers the camera when a part is in position. The image is captured under controlled light, cleaned up and passed to a model on an edge device. Rules convert the model output into pass, fail or review, and the PLC rejects bad parts while the station logs every result.
How many images are needed to train a visual inspection model?
There is no fixed number, because it depends on how much good parts vary and how different defects look. What matters more is coverage across material lots, shifts, product variants and lighting conditions. Rare defects can start with anomaly detection trained on good parts, while real defect images are collected over time.
Can AI vision detect defects it has never seen before?
Partly. Anomaly detection models learn what normal parts look like and flag anything unusual, so they can catch new defect types. They cannot name the defect, though, and they may also flag harmless changes. A common practice is to route anomalies to manual review, then add confirmed examples to the training set for classification.
What camera do I need for AI visual inspection?
Start from the smallest defect you must catch and the size of the field of view, then choose resolution so several pixels cover that defect. Moving parts need a global shutter or strobe lighting. Compare sensors on EMVA 1288 data, and remember that lighting often matters more than the camera itself.
Can machine vision inspection run without an internet connection?
Yes, and it should. Inference and the reject decision run on a smart camera or industrial PC next to the line, so a network outage never stops inspection. The cloud or data center handles training, model versions and cross-site analytics, and it syncs with the station whenever the connection returns.
How accurate is AI visual inspection?
Accuracy depends on imaging quality, data coverage and defect type, so a single headline number means little. A model can score very high accuracy while missing most real defects if defects are rare. Measure escapes and false rejects separately for each defect class, using held-out lots, before trusting any accuracy claim.
How much does a machine vision inspection system cost?
Costs vary widely from one station to the next. The budget covers cameras, optics, lighting, fixtures and edge compute, plus labeling, integration, validation, licences, maintenance and changeover work. Build the business case from your own baseline of escape costs, false-reject scrap and inspection hours rather than from generic vendor savings figures.
