Computer Vision Quality Inspection: A Practical Guide for UK Manufacturers
Customer discovers scratches on a casing which were not detected during three rounds of inspections. No one was negligent. It is natural for people to get tired after working for several hours; no one can examine all units at full speed on the production line. This is the problem that computer vision quality inspection tackles.
Proven technology, but not widely used within UK factories. This article describes how this technology can be applied to your business, what considerations are needed and how to set up a pilot. We cover data protection, compliance and funding considerations to get started. If you want an extra pair of eyes on your project, we can help you scope using our bespoke computer vision development team.
What is computer vision quality inspection?
A running solution has four components. First, cameras and lighting take a uniform picture of each unit. Second, a trained model classifies a part as good or bad. Third, decision rules map its classification into pass, fail or review. Fourth, the integration connects it to your line controls, MES, ERP or quality system. The model gets all the press, but lighting, alignment and integration make or break a project.
AI visual inspection vs traditional machine vision
Rule based machine vision systems allow specific checks to be performed on an image (edge detection, pixel intensity ranges, barcodes, for example). However, deep learning inspection allows images to be classified by training on labelled examples. This can be useful for checking surface finish, welds and cosmetic defects. In many factories, a combination of rule-based measurement and deep learning classification is used.
Common use cases
The usual tasks performed are finding cracks or scratches, ensuring that no single piece is missing in the assembly, and label verification. Label verification is especially important in the food, beverages, and drug industries since incorrect labeling poses a legal risk.
Where UK manufacturers stand today
In 2026, Make UK conducted a survey among its members and the numbers are disappointing. Just 2% claim that AI is embedded in all operations of their company. Thirty-seven percent have small pilot projects, 43% are experimenting and 18% have not even begun.
The biggest gap lies on the shop floor. Eighty-three percent of manufacturers employ AI in back office operations, including human resources, marketing, and finance. Only 11% use AI in production operations and 6% use it in quality control. It is explained by Make UK by the complexity, cost and skill required for such projects. Back office operations are less costly and more easily implementable than those which are connected to production machinery.
If you do it correctly, it gives you an opportunity since your competitors most probably have not yet automated the inspection process. The biggest barrier identified by Make UK is skill and capability, mentioned by over half of the respondents.
When is computer vision quality inspection a good fit?
It pays back fastest where inspection is a real bottleneck or a real risk. It is a strong candidate when:
- Inspection volumes are high or production lines operate at speed.
- Manual inspections are inconsistent, repetitive or difficult to perform reliably.
- Defects can be captured by a camera and clearly defined through inspection criteria.
- Identified defects lead to a practical action, such as product diversion, rework or an alert.
- The present expenses associated with defective products, scrap, and rework are clearly understood.
AI visual inspection can not be the right solution in every situation. It is less suitable when defects occur inside products and cannot be captured by a camera, or when lighting and camera positioning are difficult to control. It may also struggle when quality standards rely heavily on subjective judgement.
For low-volume production, manual inspection by a trained professional may be more cost-effective. Assessing these limitations early can prevent unnecessary investment. If the suitability is unclear, an AI consulting service can help evaluate technical feasibility and potential costs before committing to hardware.
Bonus Read: How to Hire Computer Vision Consulting Experts: Costs, Compliance & Best Practices
How to run a computer vision quality inspection pilot
Make UK suggests a “scan, pilot, scale” methodology. Identify potential use cases, test them out in controlled trials, and grow them only when they succeed. In terms of inspection, it involves six stages.
1. Define the inspection decision
Defects to be considered, severity, action required (re-work, hold, reject or accept). Method, sample rate and line speed. The system may flag items as defective, however, decisions around final disposition of items should be made as part of your quality processes. Therefore, final disposition should be taken by authorised quality personnel.
2. Standardise the scene
Uniformity is key. If your model is not trained under uniform conditions it will suffer when lighting conditions change during inspection in the plant. Ensure that the camera is in a fixed location, lighting in the inspection area is controlled, and that mechanical protection is provided for the kit in order to take training images.
3. Build a defect library
Collect images of good products and all defect types, including rare cases. Ask quality specialists to label borderline examples and document inspection standards. This creates consistent training data. Our guide to preparing business data for AI covers these data quality principles in more detail.
4. Run in shadow mode
Have the system operate side by side with your inspectors without intervening in the line operation. Compare the classifications made by the system to those made by your inspectors over a period of several weeks. Shadow mode will reveal lighting issues, lack of training data, and ambiguous defects without sacrificing any production.
5. Integrate and go live
Once the results are consistent, connect the system to your MES, ERP or quality management platform. Record each rejected item, its image and the reason for rejection against the relevant batch. Also, establish a manual inspection process to follow if the system goes offline. Our guide to integrating AI into your business explores the wider integration considerations.
6. Monitor and retrain
Products, vendors and lighting systems may all be changed and upgraded, and model drift is common. Think of the system as an ongoing managed process to measure and change, not an installation. Our AI infrastructure and MLOps practices, including versioning, monitoring, and controlled release, all apply here as well.
Measure what matters: escapes and false rejects
Every inspection regime makes two types of errors. Either there is an escape of defect or false rejection of a good item, leading to scrap and rework. Calibrate the model to capture all defects and false rejects will shoot up. Calibrate it for minimising false rejects and defects escape more often. Measure both on a single dashboard and calibrate by the cost incurred by each error.
Measure the baseline before implementing the pilot phase: the current level of escapes, scrap, rework, and inspection labor cost. Without this, you cannot measure the benefit, and without benefit, it is difficult to get the approval for implementation.
Be wary of headline accuracy claims. An often-quoted McKinsey analysis suggests AI-based visual inspection can improve defect detection by up to 90% compared with human inspection. That is an upper limit under good conditions, not a planning figure, and vendor pages often quote higher numbers with little context. Test on your own product, lighting and defect library.
Build, buy or customise?
You will have three possible options. The correct option depends on how standardised your inspection process is.
| Route | Best when | Look for |
| Off-the-shelf inspection platform | Defect types are common and you want a fast start | Per-station licence fees, vendor lock-in and where your training data is held |
| Custom development | Products are unusual, integration is complex or you want to own the IP | Needs machine learning and vision engineering capacity, and a longer route to first value |
| Hybrid | The platform takes care of capturing and inferring whereas the customised software manages integration and reporting | Ownership of each layer and the data is clearly established |
Costs depend on the number of cameras, production speed, defect complexity and integration requirements. Request a tailored quotation based on your inspection needs rather than relying on general price estimates. IIH Global estimates a typical computer vision project timeline of 6–16 weeks, depending on its complexity. For a fair comparison, provide every supplier with the same sample of real defects to evaluate their solutions.
Data Protection and Regulation
UK GDPR and Workplace Cameras
Most industrial cameras monitor products and not individuals. If identifiable individuals appear in the video footage, however, then UK GDPR and the Data Protection Act 2018 will apply.
Businesses must ensure that there is a lawful basis for the monitoring of videos, risk assessments are conducted and notices are given under the guidance of the ICO. Businesses can minimise their concerns by pointing cameras at products and masking identifiable information.
The EU AI Act and the UK Approach
Manufacturers of any entity within the European Union or who export their products into the EU need to comply with the EU AI Act. Due to the Digital Omnibus Act, the deadline for some AI systems’ compliance has been pushed back to 2027 and 2028, depending on the applications used.
It is worth noting that AI for quality control purposes is not necessarily high-risk because it is applied to a regulated product. However, there can be more stringent regulations in place when there is a possibility of endangering health or safety due to system failure.
Funding: What Made Smarter Can Cover
English SME manufacturers could get help from the Made Smarter Adoption scheme. It provides match funding grants of up to £20,000, covering up to 50% of the eligible project cost, along with digital guidance and leadership training.
This scheme is designed for manufacturers and engineering companies with 249 employees or less. However, as the funding is provided on a regional basis, there may be variations in eligibility requirements and available funding.
For instance, to get a full grant of £20,000 at the 50% funding rate would mean that the project has to cost £40,000. Projects costing less than this amount could be eligible for funding too but proportionally less. Computer vision for quality inspection could be considered an eligible project in digital technology.
Firms located outside of England should find out what regional funding schemes apply to their business.
Common Mistakes to Avoid
Vision-based systems are likely to fail due to neglecting the practical considerations when developing such applications. The most frequent problems include the following:
- From the AI model: State the inspection problem and the action needed before the technology is selected.
- Overlooking the inspection setting: Low light, camera placement and changes to the manufacturing environment may impact the results.
- Focusing solely on accuracy: Keep track of false rejects and missed defects to see how the system works.
- Missing out on shadow-mode: Use the system while the current inspection process is ongoing before rejecting the products.
- Neglecting quality teams: Get the inspectors involved from the start so that they are prepared for exceptions.
- Ignoring regular maintenance: Track performance, update the training data and retrain the model when the product changes.
Ending Note
Computer vision quality inspection is a realistic solution for manufacturers seeking enhanced accuracy in defect detection and improved consistency in production processes. To achieve this, however, one must first define a well-formulated problem statement. Create consistent conditions for testing, perform tests in shadow mode, calculate false rejects, along with missed defects, and include quality professionals in the process from day one.
Other things to consider are funding possibilities, privacy issues, and regulations. Tackling those ahead of time can mitigate the risks of deployment while making scalability of the solution much easier when the time comes.
IIH Global assists manufacturers in evaluating computer vision feasibility, creating defect datasets and implementing AI-based quality inspection solutions tailored to your production needs. Learn more about our AI computer vision development services.
Ready to improve your quality inspection process? Get in touch to discuss your production requirements and explore the right approach for your business.
Share On :