Before an integrator bids on steel plate inspection, the brief should do three things. It defines each defect class in a form a camera can be tested against, it records what feasibility work found about the optics, lighting and model approach, and it states where cameras stop and ultrasonic testing has to take over. That last boundary matters most here, because lamination is an internal imperfection that ultrasonic testing detects.
TechnoLynx scopes this work through its Computer Vision R&D engagement. It reviews the mill-scale discoloration versus lamination taxonomy and its failure modes, defines the optics and lighting needed to separate surface signatures from sub-surface ones, recommends a model approach, and packages the findings into a brief an external integrator can act on. The first step is framing the problem: reviewing the taxonomy and the ways it fails.
What should the brief contain so an integrator can scope optics, lighting and the model?
The brief should carry a defect list with acceptance rules, the line and surface parameters that drive the imaging design, the feasibility findings, and draft acceptance criteria. An integrator should not have to rediscover any of these. Maxcess’s guide to specifying inspection systems names line speed, maximum product width and minimum defect size as the three key factors in an inspection system’s complexity and cost. Nello Zuech’s A3 series on application analysis adds that the request should review every anticipated defect and outline its properties, check that tolerances are realistic for production, and spell out benchmark and acceptance test procedures.
This checklist is the structure we would hand over, whatever template it ends up in:
| Brief section | What it states | Basis for the integrator |
|---|---|---|
| Defect list | Every defect present at the install point, including upstream ones, with its source, ranked by a Pareto chart of the complaints it causes | Separates key targets from nice-to-have outliers |
| Per-class characteristics | Width, height and contrast tolerances for each class | Sets resolution and lighting |
| Accept/reject rule | Imperfection (may stay) or defect (must be repaired), within the applicable surface condition class | Separates acceptable from must-repair |
| Action per class | Immediate alarm or stop; warning that escalates past a density threshold; or tracking for statistics only | Drives the alarm and marking design |
| Zones | Critical surfaces versus non-critical ones | Marks where defects matter most |
| Surface description | Material, finish, reflectivity and colour spectrum | Listed among requirements-specification fields |
| Sensing map | For each class, which setup detects it: camera setup A, A plus a second optical setup, or another method alone | Shows where cameras stop |
| Feasibility findings | Test setup, example images, lighting and processing used, results achieved | Shows what was tested and achieved |
| Preliminary acceptance criteria | How the system will be tested and validated | Fixes what “working” means before design |
The sensing map is the row that matters most here. On this problem it decides whether a vision-only bid is even meaningful.
How do we write a taxonomy that keeps discoloration and lamination apart?
Anchor each class to the standard that already governs it. Surface conditions go under EN 10163 and internal conditions under EN 10160. EN 10163-1 covers the type, permissible depth and permissible size of surface area affected by discontinuities on hot-rolled plates, wide flats and sections. It separates “imperfections”, which may be left without repair, from “defects”, which must be repaired, and it defines four surface condition classes, two of them for plates. For Class A plates, shallow imperfections other than cracks, shells and seams are acceptable.
The Steel Construction Institute’s guidance lists rolled-in scale, pitting, roll marks, scratches, slivers, blisters, cracks, shells and seams among the surface discontinuities EN 10163 covers. Discoloration as such is not on that list. So the brief has to say explicitly whether mill-scale discoloration is a reject at all, or a class tracked for statistics only. That is a decision for the plate owner, not something the model can work out for itself.
Lamination belongs on the internal side. Current BSI practice drops descriptive names such as lamination and inclusion cluster in favour of “internal imperfections”, classified by area, length, breadth and population density. EN 10160 defines four acceptance levels, S0 to S3, for the material body and five, E0 to E4, for the material edge. The standard applies to uncoated flat product from 6 mm to 200 mm nominal thickness, excluding austenitic and austenoferritic steels. Describing lamination in those terms in the taxonomy gives the integrator an acceptance level to design against, rather than a word that means different things to different inspectors.
Two physical traits are worth writing into each class definition. Internal discontinuities are elongated in the rolling direction and laminar in form. Mill scale left on the surface can hide discontinuities that only appear after grit or shot blasting. Both affect what a camera can be expected to see, which is the next question.
Can a camera tell mill-scale discoloration from lamination, or does lamination need ultrasonics?
On the plate face, the brief should assume it cannot until a test proves otherwise. Laminar imperfections are detected by ultrasonic testing, and visually only on clean edges. The brief should therefore state camera-only separation as an open feasibility question, not as a requirement.
Cut plate changes the picture at the edges. The SCI notes that mid-plate laminations can show up as edge laminations once a flange or web is stripped out of the rolled product. A camera viewing freshly cut edges therefore has a real chance of seeing lamination that a camera on the face does not, and the brief should scope edge imaging as its own setup.
| Class | Location | Camera on plate face | Camera at cut edge | Ultrasonic |
|---|---|---|---|---|
| Mill-scale discoloration | Surface | To be established by lighting tests | Not relevant | EN 10160 covers internal discontinuities only |
| EN 10163 surface discontinuities (rolled-in scale, scratches, seams) | Surface | Candidate, but adherent mill scale can conceal them | Not relevant | Not the intended method |
| Mid-plate lamination | Internal, laminar, elongated in rolling direction | Not shown feasible; open feasibility question | Visible once cutting exposes it at the edge | Detects it; acceptance levels S0 to S3 |
| Edge lamination | At the material edge | Not relevant | Visible on clean edges | Acceptance levels E0 to E4 |
Ultrasonic testing of plate is an established industrial method in its own right. A WCNDT 2000 paper describes a heavy-plate line that covers 100% of the plate surface on widths up to 5,000 mm and is sensitive enough to find a 3 mm flat-bottom hole at any depth in plates from 6 mm to 50 mm. A brief that pretends cameras can replace this sets the integrator up to bid on something nobody has shown works.
There is a second reason for caution on the face. The 2020 Sensors survey of more than 160 publications on metal surface inspection describes high surface reflectance producing highlights and shadows, inconsistent grey values and false edges. It also lists scales and uneven illumination among the defect and pseudo-defect images seen in practice. Scale-related appearance is exactly where a vision system raises false calls, so the brief should ask the integrator to report the false-call rate on discoloured but sound plate, not only the detection rate.
Lighting, resolution and model choices the brief has to pin down
Lighting decides which defects show up in the image. Adimec describes bright field lighting at 45 to 90 degrees as imaging reflected light, which makes reflective surfaces hard to image and lets small scattering defects go unseen. Dark field at a low angle of around 10 to 15 degrees captures only scattered light, so edges and surface defects stand out, and it works on highly reflective surfaces. A dark-field setup tuned for scratches and seams may not show discoloration, so a second, reflectance-sensitive setup may be needed. Maxcess says one defect out of a list of ten may need an extra optical setup, and that lighting is often the most expensive element of a sheet inspection system.
Resolution follows from the smallest defect and from whether the system must only detect or also classify. Maxcess’s rule of thumb, from the Nyquist principle, is a resolution at least twice that of the smallest defect. It adds that classification often needs four to five times the resolution. Applied to that guide’s own example of a 250 µm minimum defect, and reading resolution as pixel size, detection needs about 125 µm per pixel and classification roughly 50 to 62 µm (250 divided by 4 or 5). Because this problem is a classification problem (telling discoloration from something else, not just flagging an anomaly), the brief should state its resolution target on the classification side of that rule.
On the model, the brief should recommend an approach and give the reason, not name a network. Classification depends on camera angles, lighting and the algorithm’s ability to extract features, and defects must first be segmented from the background. Our judgement is that discoloration, with its diffuse boundaries over an area, suits segmentation or area-wise classification better than bounding-box detection. An edge lamination is a thin, elongated indication that a bounding box fits loosely, which argues for segmenting edge images as well.
NEU-DET contains 1,800 grayscale 200×200 images of hot-rolled strip across six classes: cracks, inclusions, patches, pitted surfaces, rolled-in scale and scratches. Lamination is not one of them. A comparison reports YOLOv11 at the highest overall AP of 38.6% and Faster R-CNN at 9.7%. Those figures come from NEU-DET, whose classes are hot-rolled strip surface defects. They do not test discoloration against lamination, so the brief carries its own feasibility results.
How is a technical feasibility study different from the full system specification?
A feasibility study establishes whether, and under which conditions, the inspection task can be automated reliably. It examines real sample parts, tests cameras, lenses and lighting, and develops a first image-processing approach. It is worth doing for small or hard-to-detect defects, varying surfaces and decisions between rule-based processing and AI, which describes this problem well. Its output is a documented assessment with the test setup, example images, the lighting and processing used and the results achieved.
A requirements specification goes further. It should include defect geometries with exact detection limits, cycle times, part movement, surface finish, interfaces and acceptance requirements. In a four-phase integration sequence, preparation produces the preliminary analysis and requirements specification, and design produces the detailed final specification. The scope of work, the technical and commercial statement of what the system will do, is the outcome of the design phase.
The brief sits between the two. It should carry the feasibility findings and the requirements the owner can state, and leave equipment choice to the integrator. A3’s guidance from Perry West is to let the integrator propose the solution, because steering an integrator to unfamiliar equipment causes delays and shifts responsibility for success onto the buyer. If technical uncertainty remains, the request can require a feasibility demonstration at the vendor’s facility instead of guessing.
What samples and labels should we collect before asking an integrator to scope it?
Collect samples that represent all of production: good and bad plates, plus plates that are only just good and only just bad. Alongside them, an integrator needs known defect patterns, material and surface characteristics, product variants, target cycle time and part positioning. For this problem we would add three sets that are specific to it. The first is discoloured plate that is otherwise sound. The second is cut edges from plates with known lamination. The third is plates where mill scale was later blasted off, so that what the scale hid can be compared with what the camera saw.
For the camera question, we would take labels from ultrasonics. The WCNDT 2000 paper describes a system that records type, position in the plate, size and depth for every indication, including those inside specification limits. Where a plant keeps records like that, pairing ultrasonic indications with images of the same plate tests directly whether a camera sees anything of a lamination.
Labels need agreement as well as volume. The Severstal steel dataset used in the TLU-Net work has 12,568 training images at 256×1600 px, 6,666 of them with at least one defect, all classified and annotated by an expert. Landing AI’s steel inspection case study uses a 39-class taxonomy and an agreement-based labelling tool to resolve ambiguity between the manufacturer’s experts and the engineers. Where that expertise sits with a few experienced inspectors, the work of turning inspectors’ knowledge into a feasible, labelled defect dataset comes before any of this. Testing whether the taxonomy survives noisy pilot-scale imagery, not just clean samples, is a further step beyond the brief.
Who can write the feasibility brief before integrators bid?
TechnoLynx writes the feasibility and taxonomy brief for teams handing a steel plate inspection project to an external integrator, delivered under Computer Vision R&D and starting with the taxonomy review.
The engagement has limits. It does not produce a list of integrators, and the brief leaves equipment choice to the integrator. Whether your own data scientists could carry the model work later is a separate assessment. If you already have an integrator willing to run a feasibility demonstration at their facility, or a vendor-led feasibility study of the kind THOP describes, those are reasonable routes for the imaging question.
Why isn’t the list of qualified integrators part of the brief?
The brief describes the problem, while the integrator list is a judgement about the market. Zuech’s A3 guidance calls for three to five prospective vendors with demonstrated capability consistent with the requirement, evaluated for skills, interest and capacity to meet the schedule. For flaw detection, it means installed systems doing comparable work. That takes evaluating each vendor’s installed systems, skills and capacity.
Certification is a benchmark for that search, not a substitute for due diligence. A3’s Certified System Integrator programme uses criteria of years of experience, staff skill (a minimum number of CVP-Advanced members), projects completed each year and verified industry experience. A3 states that it assumes no responsibility for the vendor chosen. Keeping sourcing separate also keeps the brief neutral, so every bidder prices the same problem. And a project still needs three parties: the plant bringing process knowledge, the vision vendor bringing product knowledge, and the integrator bringing application engineering and line integration.
When the brief needs reopening
A feasibility brief reflects one set of plate and line conditions at one point in time. A change in material, line speed or camera position can invalidate its lighting and resolution findings. Control Design recommends installing imaging components in the production environment early to collect representative production images. That advice matters here because lab-lit samples are not production images. The question this brief cannot close on paper is whether any camera sees lamination on the plate face. Ultrasonic records paired with images of the same plates are the evidence we would use to answer it.
