Why Your Camera System Is Only as Good as the Software Behind It
CTO, Deep Purple AI Consulting
Machine vision software is the processing layer that sits between an industrial camera and a quality decision. It handles image pre-processing, defect classification, reject logic, traceability, and reporting. In Irish manufacturing, that typically means label checks, barcode verification, seal integrity, fill level, and track-and-trace on regulated lines.
Most Irish manufacturers who invest in machine vision start with the hardware. A Keyence line scanner. A Cognex smart camera. A Basler industrial camera. The sales engineer runs a demo. The purchase order goes through.
Six months later, the system catches some defects but misses others. False reject rates are too high. Good product is being binned. Operators override the system when it flags too many false positives. Batch release is delayed by inspection exceptions that nobody can explain from the data.
This article is for manufacturers with 10 to 500 employees who already have cameras and need better software. Not for multinationals running 24/7 lines with dedicated automation engineering teams.
The camera is not the problem. The software is.
Why Most Vision Inspection Projects Disappoint
There are three reasons machine vision systems underperform, and none of them are about the camera.
Lighting. Industrial cameras need consistent lighting to produce consistent results. A system that works on a Tuesday morning may struggle on a Thursday afternoon when sunlight hits the line differently. If the software cannot handle that variation, it will never be reliable.
Integration. The camera captures an image. But what happens next? Does the software talk to the PLC? Does it log to your MES, SCADA, or ERP? Most off-the-shelf vision systems handle the image processing but leave integration to you. That is where projects stall.
Training data. A vision system is only as accurate as the data it was trained on. Fifty images of "good" and fifty of "bad" will work in a demo. It will not work on a production line where defects come in hundreds of variations. Building a dataset that reflects real production takes time and iteration.
What Machine Vision Software Actually Has to Do
Machine vision hardware captures images. The software layer turns those images into decisions. It must handle five things within the cycle time of your line. At 200 units per minute, that is 300 milliseconds per image.
Pre-processing. Normalising lighting, removing background noise, aligning images for comparison.
Classification. Deciding whether a product passes or fails. This can be rule-based (pixel thresholds) or ML-based (trained models). Most production systems use both.
Reject logic. What happens when a defect is detected? Trigger an air jet? Stop the line? Log and continue? The answer depends on the defect type and the cost of a false reject versus a missed defect.
Traceability. Every inspection result logged with timestamp, batch ID, image, and classification. This is not optional for regulated manufacturing.
Reporting. Shift summaries, defect trend analysis, operator dashboards. The data a vision system generates is often more valuable than the reject decision itself. See our data analytics services for how we turn inspection data into operational insight.
If any of these layers is weak, the whole system underperforms.
The Training Data Advantage
This is where custom software separates from off-the-shelf.
Off-the-shelf vision software ships with a fixed model or a small set of configurable rules. It does not learn from your production line. Custom software does. Every image your system processes is training data for the next version of the model. In my experience, the training data conversation is where most projects succeed or fail.
We train models in PyTorch and deploy using ONNX Runtime for edge inference. The model runs locally on your hardware without sending images to the cloud. When we retrain, we pull from your accumulated production dataset. After six months, a system that was 92% accurate at launch might be 97% accurate. After a year, your operators are only reviewing the genuinely ambiguous cases.
The system gets better because it is learning from your products, your defect types, your edge cases. Not from a generic training set built by a camera vendor.
Working with Your Existing Hardware
You do not need to replace your cameras. Custom machine vision software works with the hardware you already have.
Keyence, Cognex, Basler, or any industrial camera with a standard interface. GigE Vision, USB3 Vision, and Camera Link are all supported. We do not sell hardware. We build the software layer that makes your existing hardware more useful.
If you run multiple SKUs on the same line, the software needs to handle changeover without retraining every time. That is a software architecture decision, not a camera decision.
If your hardware vendor recommended you speak to us, that is because we build the software layer they do not.
If you already have cameras and want to know whether custom software would improve them, book a 20-minute feasibility call.
Already have cameras on the line? Let's look at whether custom software would improve what they catch.
Book a 20-Minute Feasibility CallWhere a Hardware Integrator Ends and a Software Platform Begins
For a single standard vision inspection station, a hardware integrator may be the right choice. They handle camera selection, lighting, mounting, and station programming. These are not competing approaches. In many projects, the best outcome is an integrator handling the hardware and a software partner handling everything above the camera.
Custom software starts to make sense when the job goes beyond the station:
| Hardware Integrator | Software Platform |
|---|---|
| Camera selection and mounting | MES, SCADA, and ERP integration |
| Lighting design | Cross-line and cross-site reporting |
| Station-level reject logic | Model retraining and improvement |
| Single inspection point | Multi-station pipeline |
| Vendor-specific tools | Audit trails and role-based access |
| Data ownership (software and models) |
Validation and Traceability for Regulated Manufacturing
If you manufacture food, pharma, or medical devices in Ireland, your vision inspection system must meet specific standards. GAMP5 for computerised systems. 21 CFR Part 11 for electronic records. FSSC 22000 or BRC for food safety.
In practice, this means: every inspection result stored with a complete audit trail. Access controls on classification thresholds. Versioned models with documented change management. Validation protocols (IQ, OQ, PQ) built into the software design, not bolted on. Batch release evidence generated automatically.
Most off-the-shelf machine vision software does not handle this compliance layer. For pharma and medtech, validated machine vision systems are a regulatory requirement. Custom software builds compliance into the architecture from day one.
Which Inspection Tasks Are Worth Automating
Different products need different inspection logic. Here are the common machine vision inspection tasks in Irish manufacturing:
| Inspection Task | What the Software Checks | Common Industries |
|---|---|---|
| Label verification | Correct label, position, orientation, batch code (OCR) | Food, pharma, medtech |
| Seal inspection | Seal present, continuous, correctly positioned | Food packaging, pharma blister packs |
| Cap inspection | Cap present, straight, correctly seated | Beverages, pharma, cosmetics |
| Fill level | Product filled to correct level | Food, beverages, pharma |
| Surface defect detection | Scratches, dents, discolouration, contamination | General manufacturing, medtech |
| Barcode and QR verification | Code present, readable, correct for batch | All regulated manufacturing |
Manual visual inspection accuracy typically ranges from 70% to 85%, depending on complexity and fatigue. Automated vision inspection consistently achieves 95% or above on the same tasks.
When Computer Vision Is a Waste of Money
Not every inspection problem needs computer vision. Here is when it does not make sense:
- Low volume. If you inspect 50 units a day, the ROI will not justify the investment.
- Highly variable products. If every product looks completely different, training a vision model is impractical. CV works best when the "good" product is visually consistent.
- No clear definition of "defect." If your QA team cannot agree on what a defect looks like, a vision system will just automate the disagreement.
- The problem is upstream. If defects are caused by a process issue (wrong temperature, worn tooling), catching them at inspection is treating the symptom. Fix the process first.
If any of these apply, we will tell you. There is no point building a system that will not deliver value.
What Usually Breaks First
Three things go wrong in the first month of most vision deployments.
Lighting drift across shifts. The system was tuned during the day shift. The night shift has different ambient light. Accuracy drops and nobody notices until reject rates spike.
New SKU or packaging artwork. A new label design or product variant arrives. The model has never seen it. It starts rejecting good product.
Reject logic misaligned with operators. The system flags a defect but the operator cannot see it, does not trust the flag, and overrides it. After a week, the override rate is 40%.
Custom software handles all three. But you should know they are coming.
What a Safe Pilot Looks Like
A pilot runs for 4 to 6 weeks and answers one question: can we reliably detect the defect types that matter most to your business?
What you get:
- A trained model for your top 2 to 3 defect types, using your production images.
- Integration with your existing camera hardware.
- A dashboard showing inspection results, defect rates, and false reject rates.
- Baseline comparison: manual reject rate versus automated reject rate.
Acceptance criteria (agreed before the pilot starts):
- Target detection rate for each defect type.
- Maximum acceptable false reject rate.
- Operator review workflow for borderline cases.
- What happens if targets are not met.
What it costs: Typically €15,000 to €25,000. This is often 50% grant-funded through Enterprise Ireland or LEO programmes. Your net cost could be €7,500 to €12,500.
What happens after: If the pilot works, you have a specification for the full system. A single-line deployment typically takes 3 to 6 months from pilot sign-off. If it does not work, you know why, and you have not committed to a six-figure project to find out.
Deep Purple has 13 years of production software delivery behind it. We have deployed production computer vision on Irish construction sites and delivered predictive quality modelling for an Irish food manufacturer. The manufacturing proof here is adjacent, not identical. That is why the right first step is a pilot. You own everything we build: the software, the trained models, and the data.
Book a 20-minute feasibility call. We will look at your inspection task, line speed, camera setup, and compliance requirements. If a standard integrator is the right answer, we will tell you.
How to Tell If This Is Worth a Feasibility Call
If you already have cameras on your line and they are not performing as expected, the issue is almost certainly software. Bring these to the call:
- Your current inspection problem and defect types.
- Sample defect images if you have them.
- Line speed and throughput requirements.
- Existing camera hardware (make, model, interface).
- Systems to integrate with (MES, ERP, SCADA).
Not sure if your process is ready for computer vision? Use our CV Readiness Checklist.
For a full overview of how we approach computer vision projects, see our computer vision service page.
Deep Purple builds custom machine vision software for Irish manufacturers that works with existing Keyence, Cognex, and Basler cameras, providing automated visual inspection, defect detection, traceability, and compliance-ready audit trails for regulated production environments including food, pharma, and medical devices.
Frequently Asked Questions
Start with a 20-Minute Feasibility Call
We will look at your inspection task, line speed, camera setup, and compliance requirements. If a standard integrator is the right answer, we will tell you.
Book a 20-Minute Feasibility CallOr start with a €1,250 AI Assessment →Related Reading
About Barry Gough
CTO, Deep Purple AI Consulting
Barry Gough is the CTO of Deep Purple AI Consulting. With an MSc in Computer Science from University College Dublin, where machine learning was a core focus of his studies, and over 20 years building production software systems, Barry brings formal ML training and deep hands-on engineering experience to every AI and data analytics engagement.
Barry completed his masters at UCD in 2011, studying ML algorithms, statistical modelling and data-driven systems just as big data techniques were maturing and deep learning was about to transform the industry. At Purpledecks (Deep Purple's predecessor consultancy), he spent nearly a decade progressing from Senior Developer to Head of Operations, leading the technical delivery of enterprise projects that increasingly incorporated machine learning, computer vision, data classification, predictive features and recommendation engines for commercial clients across Ireland and the UK.
In 2023, as CTO of Reactable AI, Barry architected and built an autonomous AI marketing engine from the ground up, a self-learning system that generates and optimises marketing campaigns across channels. This was one of Ireland's earliest production deployments of autonomous AI agents, requiring him to design systems where AI made real decisions with real consequences.
At Deep Purple, Barry leads all technical delivery: AI system architecture, machine learning model development, data pipeline engineering, and manages a team of experienced ML engineers and applied statisticians. His combination of formal ML education, a decade of incorporating AI into commercial projects and hands-on experience architecting autonomous AI systems means clients work with a technical lead who can make genuine engineering decisions about AI.
Deep Purple AI Consulting (deeppurple.ai) is an AI consultancy and custom software development company based in Ireland. We help established businesses identify where AI can make a real difference, then build the systems to make it happen. Senior-only delivery. Grant-funded where possible. No hype.