Opsio - Cloud and AI Solutions
Cold-rolled steel strip passing between a sensor housing above and a second housing below it on a portal frame, with a control cabinet beside the line and coils in the mill hall behind

AI Visual Inspection for Metals & Materials

Find scratches and inclusions on metal strip, seeds and stones in glass, and holes and streaks in paper, textile and film webs, on every metre at full line speed.

Book a walkthroughSee use cases

What is AI visual inspection for metals and materials?

AI visual inspection for metals and materials is a camera-based surface inspection system for continuous strip, glass and web. Opsio builds it with line-scan cameras, bright-field and dark-field lighting and deep-learning models that find strip surface defects, glass body and surface defects, and web defects, and map each one to its coil or roll position.

Key facts

  • Line-scan cameras above and below the strip image both surfaces at hundreds of metres per minute.
  • An encoder on a line roller gives every defect a position on the coil or roll.
  • Inference runs on edge hardware beside the line in under 50 ms per image tile.
  • Defect maps go to Level 2, MES and slitters over OPC UA, PROFINET or TCP/IP.
  • Images and coil maps stay on-premises by default; optional retraining runs in an EU cloud region.
  • A feasibility study on your samples typically takes 1 to 2 weeks in the Opsio AI Lab.

Main defect families

Strip surface defects
Scratches, inclusions, roll marks, scale, slivers and coating faults on steel, aluminium or copper strip.
Glass body and surface defects
Seeds, stones, cords and tin defects in float glass, plus chips and checks on glass containers.
Web defects
Holes, streaks, weave faults, stains, wrinkles and gels in paper, textiles, nonwovens and film.

Last reviewed by Praveena Shenoy, Head of AI & Data, Opsio, a Swedish cloud and AI services company. Part of Opsio's AI Visual Inspection service.

Why is surface inspection hard on strip, glass and web lines?

Your material never stops, so every square metre has to be judged in the moment it passes a camera line. Cold-rolled strip at 600 m/min moves 10 metres every second, up to about 2 metres wide, with two surfaces to check.

The surface changes more than the defects do. Bright annealed strip mirrors the light, hot-rolled strip carries scale and steam, galvanised coating has spangle, and textile structure hides a slub. A fixed threshold either floods the operator with false rejects or lets a sliver through.

Human inspection breaks down because a person sees one surface, at an inspection stand, after the line has slowed. Many defects repeat: a damaged work roll marks every revolution until it is changed, so a late find means kilometres of affected product.

The cost of an escape lands downstream. A coil with an undetected sliver gets slit, shipped and pressed into panels before the claim arrives; a float sheet with a stone can break in tempering; a gel in film can puncture during converting.

Steel, glass and web defect detection use cases

Each use case covers one production stage: your line today, the defects, the camera and lighting setup, and what changes. Webs use the same full-width line-scan approach as print and converting webs.

Steel, aluminium and copper strip

Close-up of bright aluminium strip curving over a roll, with a fine longitudinal scratch, a small pit and a raised inclusion particle on the brushed surface.
Aluminium strip with a longitudinal scratch, a pit and an inclusion.

Situation

Your strip runs through a pickling, cold-rolling, annealing, galvanising or coating line at hundreds of metres per minute, typically 0.2 to 3 mm thick and up to about 2 metres wide. Today an operator checks the top surface at an inspection stand at reduced speed and sees the bottom only at coil ends. Many surface findings arrive as customer claims.

Defects

  • ScratchesLinear grooves from strip rubbing on stationary guides, rolls that stop turning or coil handling.
  • Inclusionsnon-metallic particles rolled into the surfaceSlag or refractory particles from steelmaking that roll out into dark elongated marks.
  • Roll marksRepeating imprints from a damaged or dirty work roll, spaced at the roll circumference.
  • Edge cracksShort cracks at the strip edge from poor edge condition or heavy reduction during rolling.
  • Scaleiron oxide layerRolled-in oxide that pickling did not remove, seen as grey patches or streaks.
  • PitsSmall cavities left where scale or particles were pressed in and later fell out.
  • Sliversthin flaps of metalLaps of metal partly attached to the surface, usually from casting or slab surface defects.
  • Oil spotsStains from excess rolling emulsion or hydraulic drips, which disturb later coating and painting.
  • Coil-break linestransverse kink linesFine lines across the strip where the coil bent unevenly while uncoiling.
  • Coating defectsBare spots, dross pick-up, zinc runs and uneven spangle on galvanised or coated strip, from bath or roll conditions.

Inspection setup

You get two stations, one above and one below the strip, after the last process section and before the recoiler or shear. Each uses two to four line-scan cameras (typically 8k pixels, GigE Vision or CoaXPress) with fixed-focal-length lenses, at typically 0.1 to 0.5 mm per pixel. A bright-field LED line light sits at the specular angle and a dark-field light off-angle, strobed on alternate lines so each camera records both channels.

An encoder on a line roller triggers every scan line, so each defect gets a coil ID, a distance from the head end and a position across the width. The model classifies each finding against your surface-defect catalogue, then logs it, flags it to the operator or fires an ink marker for heavy defects. Background reading: metal surface defect detection with deep learning.

Outcome

You see both surfaces of every coil, defect by defect, while the coil is still on the line. Roll marks show up as periodic patterns on the coil map, so the crew can change the roll within one coil. The map follows the coil to slitting and cut-to-length, where flagged sections are cut out, and defect counts feed grade assignment.

Glass inspection: float and container

Clear float glass sheet on a white light table in transmitted light: a small dark stone with a distortion halo, an elongated gas seed and a chipped edge
Stone inclusion, seed and edge chip in float glass under transmitted light

Situation

A float line produces a ribbon about 3.2 metres wide that is cut into sheets after the annealing lehr, typically at under 25 m/min. Container plants form bottles and jars on IS machines (individual section machines) at typically 300 to 600 containers per minute. Many lines already have inspection; the gap is classification, because seeds, stones and surface dirt all look like dark spots to a threshold.

Defects

  • Bubblessmall ones are called seedsGas trapped in the glass when fining of the melt is incomplete, round or stretched along the draw.
  • Stonessolid inclusionsUnmelted batch, refractory or devitrified particles, often with a stress halo that can break the glass in tempering.
  • Cordsstreaks of glass with a different refractive indexPoorly mixed glass that shows as wavy optical distortion.
  • ScratchesSurface marks from rollers, handling or cullet on the conveyor.
  • InclusionsForeign particles such as metal or refractory fragments trapped in the glass body.
  • Tin defectsTin speck, tin pick-up and bloom on the bottom (tin-side) surface from contact with the molten tin bath.
  • ChipsFragments missing from a sheet edge or a container finish, from cutting, breaking-off or handling.
  • Checkssmall cracks in container glassShort cracks in the finish, neck or base from thermal shock or mould contact.

Inspection setup

On float glass, line-scan cameras above the ribbon after the lehr look at a transmitted LED line light below the glass at typically around 0.1 mm per pixel, with dark-field channels for scratches and a reflected stripe pattern for distortion and cords. Each spot fault is classified, sized and tagged with x/y position for the cutting optimiser, which cuts around it or downgrades the sheet.

On containers, a thermal camera at the hot end reports faults per IS section and mould. At the cold end, area-scan cameras with diffuse backlight check the sidewall, telecentric optics check finish and base, and polarised light makes checks flash; a pusher or air jet rejects the container. Background reading: glass bottle defect detection with vision AI.

Outcome

You get every spot fault classified, sized and positioned before the glass is cut or the container is packed. Seeds and stones are separated from surface dirt, so the cutting optimiser removes real faults and fewer good sheets are downgraded. On container lines, hot-end data per section and mould lets your forming team correct a mould without waiting for the cold-end report.

Paper, textiles, nonwovens and films

White nonwoven web passing over a steel roller with a small round hole and a dark grey contamination streak running in the machine direction.
Nonwoven web with a hole and a contamination streak.

Situation

Web lines run wide and fast: paper machines at typically several hundred to over 1,000 m/min, nonwoven and film lines at typically 100 to 600 m/min, fabric frames at typically 20 to 100 m/min. Fabric is often graded by an inspector at a backlit frame; paper and film rely on threshold systems that list unclassified spots.

Defects

  • HolesOpenings through the web from formation faults, sparks, or gels that burst during film stretching.
  • StreaksMachine-direction lines from a damaged doctor blade, a dirty die lip or uneven coating.
  • Weave and knit faultsBroken ends, missing picks, dropped stitches and needle lines from yarn breaks or worn machine parts.
  • Slubsthick places in the yarnShort thick yarn sections from spinning faults that show as raised spots in the fabric.
  • StainsOil, rust or water marks from machine lubrication, guide rolls or handling.
  • ContaminationForeign fibres, insects, dirt or degraded polymer specks embedded in the web.
  • WrinklesCreases from uneven web tension or roll alignment that become permanent in calendering or winding.
  • Gauge bandsmachine-direction thickness lanes in filmThicker or thinner lanes caused by die-bolt settings or temperature differences across the die.
  • Gelsunmelted or cross-linked polymer points in filmSmall lens-like spots from degraded or unmelted resin that distort light and weaken the film.

Inspection setup

Line-scan cameras span the full width at typically 0.1 to 1 mm per pixel, with an LED line light below the web for transmitted light (holes, gels, thin spots) and one above for reflected light (stains, streaks, wrinkles). Fabric frames add low-angle light to raise slubs and weave faults. The station sits after the last forming or coating step and before the winder, encoder-synchronised so each defect has a roll position.

The system logs every defect on a roll map, triggers a flag inserter at the winder, and passes the map to the slitter-rewinder, which stops at each defect. Gauge-band alarms are cross-checked against your thickness gauge. Background reading: fabric defect detection with deep learning.

Outcome

You get a classified roll map instead of a list of light and dark spots. Fabric rolls are graded with penalty points per defect, so the inspector checks flagged positions rather than every metre. Paper and film rolls arrive at the slitter with a map that says where to stop, and defects are removed before your customer sees them.

How a surface inspection system works on your line

The same six steps apply to strip, glass and web; the camera layout, lighting and decision point change.

Two grey camera housings and two LED line lights mounted on a steel beam above a moving metal strip, with a rotary encoder fitted to the roller shaft in the foreground.
Camera housings, LED line lights and roller encoder above a moving metal strip.
  1. 1

    Trigger and part presentation

    An encoder on a line roller triggers each scan line, so resolution stays constant as speed changes. Level 2 sends coil or roll ID, width, grade and speed for each product; on container lines a photo-eye triggers each container.

  2. 2

    Image acquisition

    Line-scan cameras (GigE Vision or CoaXPress) with fixed lenses sit above and below the material, with bright-field, dark-field or transmitted LED line lights, in air-purged enclosures that are water-cooled where heat or steam is present.

  3. 3

    Edge inference and latency budget

    Image tiles are analysed on an edge server with NVIDIA GPUs (TensorRT) or Intel OpenVINO in under 50 ms per tile. At 300 m/min, a marker 10 metres downstream leaves a 2-second budget.

  4. 4

    Decision and actuation

    Results go to the line PLC over PROFINET, EtherNet/IP, OPC UA or Modbus TCP. The PLC fires an ink marker, flag inserter or container reject, or slows the line, while the HMI shows the live defect map.

  5. 5

    Traceability and data retention

    Every defect is stored with coil or roll ID, position, class, severity and image. Maps go to Level 2 and MES; images stay on the edge or on-premises server for the retention period you set.

  6. 6

    Active learning for new defects and products

    Uncertain detections and new defect types are queued for operator review on the HMI. Confirmed labels retrain the model on-premises or in an EU cloud region with SageMaker, and it goes live after validation on recorded coils.

Which imaging finds which defect on strip, glass and web?

Choose the light that gives each defect the strongest contrast; the model then classifies it. Container lines share the backlit checks used for bottle inspection in food and beverage.

Defect familyBest-suited imagingWhy
Scratches, slivers, edge cracks and coil-break linesLine-scan with dark-field LED line light; backlight at the strip edgesGrooves and raised edges scatter light into an off-specular camera; edge cracks show against a backlight.
Scale, oil spots and coating defectsLine-scan with bright-field (specular) LED line lightReflectivity changes show as dark areas on the mirror-like strip.
Roll marks, pits and inclusionsBright-field plus dark-field, encoder-synchronisedBoth channels show the mark; in steel defect detection, the repeat distance identifies the roll.
Glass defects: seeds, stones, inclusions and tin defectsTransmitted bright-field line-scan plus dark-fieldBubbles and particles block or bend transmitted light; scatter separates bubbles from stones.
Cords and optical distortionReflected or transmitted stripe pattern (moiré)Refractive-index changes bend the stripes, which the camera measures.
Checks and chips in container glassPolarised or specular light from several angles; telecentric optics on finish and baseCrack faces reflect only at certain angles, so they flash in one view.
Holes and gels in filmTransmitted-light line-scanHoles pass light and gels act as tiny lenses against a uniform backlight.
Fabric defects: weave faults, knit faults and slubsLine-scan with low-angle reflected light and backlightLow-angle light raises thick places; backlight shows missing ends and picks.
Streaks, stains, contamination, wrinkles and gauge bandsReflected and transmitted line-scan, checked against the thickness gaugeSurface faults change reflection; gauge bands show as lanes of different transmission.

Integration with Level 2, MES and coil certificates

The system writes into the systems that already track, grade and certify your coils and rolls; your quality team keeps the release decision.

Level 2 automation
Reads coil or roll ID, product, width, thickness and speed from Level 2 and returns the defect list and coil map per coil over OPC UA or TCP/IP.
Coil and roll defect maps
Passes position-tagged defect maps to slitting, cut-to-length, glass cutting optimisers and slitter-rewinders, which cut out or stop at each defect.
Grade assignment and surface classes
Applies your rules per product and customer, such as automotive exposed-surface classes, EN 572-2 fault limits for float glass or 4-point fabric grading, and proposes a grade.
Quality certificates per coil
Supplies surface inspection results and images per coil for your EN 10204 type 3.1 inspection certificate, which your quality department issues and signs.
MES, QMS and SPC
Writes defect counts per class, shift and line section to MES and SPC, and links customer claims in your QMS to the stored coil map.
Where data lives
Images and maps stay on the edge server and on-premises storage by default; retraining runs on-premises or in an EU cloud region.

From first test to your line in five steps

The samples, defect classes and line interfaces below are specific to strip, glass and web.

  1. 1

    Opsio runs tests in the AI Lab

    You ship cut samples rather than whole coils: typically 30 to 50 defect-free sheets or roll sections and 20 to 50 samples per defect class in your surface catalogue, such as scratches, slivers, roll marks, seeds, stones, holes and gels. Opsio images them on a line-scan rig with candidate lighting, trains a first model and measures detection and false-reject rates per class.

    You provide
    Cut samples per defect class, your defect catalogue, line speed and width.
    Opsio delivers
    Imaging trials, a first model and per-class detection and false-reject results.
  2. 2

    Present results and pricing

    You see per-class results, example images from the bright-field, dark-field and transmitted channels, the recommended camera count per side and lighting, and the encoder and marker layout for your line speed and width. You then get a fixed quote for the production stations and the managed service.

    You provide
    Line drawings, mounting space, Level 2 and PLC interface details.
    Opsio delivers
    Results report, bill of materials, cycle-time fit and fixed quote.
  3. 3

    Decision and PO/MSA signing

    You decide whether to go ahead. The purchase order and master services agreement cover scope, acceptance criteria per defect class, coil and roll map formats for Level 2 and downstream lines, data ownership and support levels for the installed stations.

    You provide
    Purchase order, acceptance criteria and data policies.
    Opsio delivers
    Signed MSA, project plan and interface specification.
  4. 4

    Installation in the Opsio AI Lab

    Opsio builds the production stations on the real cameras, lighting, enclosures, encoder input and edge server, runs them on sample material and recorded images, and tests the Level 2 and PLC messages against a simulator. Acceptance tests against your criteria happen here, before anything reaches your line.

    You provide
    Interface test data and FAT attendance.
    Opsio delivers
    Pre-commissioned stations and a factory acceptance test report.
  5. 5

    Installation at the customer location

    The stations go in during a planned stop, above and below the strip or web, and are connected to the encoder, PLC and Level 2. The system runs in shadow mode next to your current inspection for an agreed number of coils or rolls, then switches to live marking and grading; operator training and the active-learning loop start.

    You provide
    A planned stop, mounting points and operator time.
    Opsio delivers
    Installation, integration, shadow-mode report, training and handover.

Pricing for the feasibility study, production stations and the managed service is on the AI Visual Inspection service page.

AI visual inspection for Metals & Materials: questions

What are the different types of glass defects?

Glass defects fall into body defects, surface defects and edge or finish defects. Body defects are seeds (gas bubbles), stones (solid inclusions), cords (poorly mixed glass) and foreign inclusions. Surface defects include scratches and tin defects on the bottom side of float glass. Edge and finish defects include chips on sheets and checks (small cracks) on container finishes. Each family needs its own light: transmitted for body defects, dark-field for scratches and polarised or specular light for checks.

What are some common defects in fabric?

The most common fabric defects are broken ends, missing picks, slubs, holes, stains and knit faults such as dropped stitches and needle lines. Woven fabrics also show reed marks and barré, while nonwovens show thin spots and contamination. An AI model trained on your fabric learns these classes on your actual weave, colour and texture, which is where fixed-threshold systems struggle.

What is a barré defect in fabric?

Barré is a repeating band of different shade or lustre across a knitted or woven fabric. It comes from yarn variations, uneven tension or mixed yarn lots on the machine. Because barré is a gradual shade change over a wide area, a line-scan camera with even reflected light shows it more consistently than a person at a frame, and the system reports its repeat distance so you can trace the feeder or yarn lot.

What paper defects can a web inspection system detect?

A web inspection system detects holes, spots, streaks, wrinkles, dirt, slime spots, edge cracks and coating defects on paper and board. Transmitted light shows holes and thin spots; reflected light shows streaks, dirt and coating faults. The AI model classifies each one, so your slitter operators and quality team act on classified defects with roll positions instead of a list of light and dark spots.

How does online steel defect detection work?

Online steel defect detection images the moving strip with line-scan cameras above and below, synchronised to an encoder, and classifies each defect while the coil is still on the line. Bright-field and dark-field channels show different defect types, the model assigns each one to your defect catalogue, and the result is a coil map with position, class and severity sent to Level 2 and the downstream slitting line.

What does a surface inspection system in machine vision consist of?

A surface inspection system consists of line-scan cameras, LED line lights, an encoder, an edge computer and software that classifies, maps and reports defects. On a strip or web line there is usually one camera bar per side, in enclosures that handle dust, heat or mist. The software connects to your PLC for marking and to Level 2 or MES for coil and roll maps.

How do you check fabric quality?

You check fabric quality by inspecting every metre for defects and grading each roll by penalty points. On an automated line, cameras scan the full width with reflected and transmitted light at typically 20 to 100 m/min, the model classifies each defect and assigns points, and the inspector reviews only flagged positions. Each roll then gets a grade and a defect map for cutting.

What does an automated fabric inspection machine do?

An automated fabric inspection machine runs the fabric past a camera and lighting frame, detects and classifies defects, and grades each roll without an inspector watching every metre. With AI models it handles patterned, coloured and textured fabrics that defeat simple thresholds. It outputs a roll map that cutting rooms use to place patterns around defects.

Can you train on a public glass or fabric defect dataset?

Public datasets help with research but are not enough for your line. Your glass composition, fabric construction, lighting and camera geometry differ from any public dataset, so production models are trained on images of your own material. Public data can pre-train a model; the feasibility study then uses typically 20 to 50 samples per defect class plus recorded images from your current system.

What software does a surface inspection system use?

A surface inspection system uses acquisition software for the cameras and encoder, a deep-learning model for detection and classification, and a defect database with viewers for coil and roll maps. Opsio runs the models at the edge with NVIDIA TensorRT or Intel OpenVINO, stores results on-premises, and exposes them to Level 2, MES and SPC over OPC UA, PROFINET or TCP/IP.

What belongs on a fabric inspection checklist?

A fabric inspection checklist covers defect types and penalty points, inspected length, width and shade checks, and the roll grading rule. Many mills use the 4-point system described in ASTM D5430, where each defect scores one to four points by length and the roll is graded on points per 100 square yards. An automated system applies the same checklist to every roll and records each scored defect with its position.

Book a 20-minute walkthrough for metals & materials

Pick the defects that cost you most, and we will show you live how a vision model finds them on a line like yours.

  • Live on your defect type, no slides
  • A reply within one business day
  • No obligation
Your call is with
  • Vaishnavi ShreeDirector & MLOps Lead
  • Praveena ShenoyCountry Manager, India

We use your details to answer your request, as described in our Privacy Policy.

Other industries