Weld inspection is a hostile imaging problem wearing a defect-detection label. The seam is bright, reflective and sometimes still glowing, the defects are safety-relevant, and unless the imaging tames that scene, no model can tell a sound bead from a flawed one.
DomainIndustrial AI, weld inspection
PlatformsFaststream Vision
ScopeSeam imaging to graded verdict
Binding constraintImage a glowing, specular seam before judging it
DisclosureRepresentative programme; customer not named
CONTEXT
Where this started.
Inspecting a weld seam inline — for porosity, undercut, spatter, incomplete fusion, bead geometry — is safety-relevant work where a missed defect can matter structurally. But before any of that, there is the imaging: a weld is bright, highly reflective and sometimes still hot and glowing, which is one of the more hostile scenes in machine vision.
Get the imaging wrong and there is nothing to inspect: glare washes out the surface, the glow saturates the sensor, and bead geometry — which is three-dimensional — is invisible to a flat image. So structured light or profiling is often needed to see the shape, and the capture has to be timed and filtered against the weld's own light.
So the system is built imaging-first: taming the specular, glowing seam into a usable image and profile, then classifying defects and grading bead geometry, on a robotic line moving along the seam.
CHALLENGES
4 problems, named.
Stated before any of them had an answer.
01
The seam is a hostile scene
Bright, reflective and sometimes glowing, a weld saturates and glares; taming it into a usable image is the first and hardest problem.
02
Geometry is three-dimensional
Bead height, undercut and fusion are shape, not just appearance; a flat image misses them, so profiling or structured light is needed.
03
Defects are safety-relevant
Porosity and incomplete fusion can matter structurally; the inspection has to be trustworthy, not indicative.
04
It moves on a robot
The inspection travels along the seam on a robotic line; capture, profiling and decision have to keep pace with the motion.
ARCHITECTURE
How it was built.
A weld that looks perfect can be unsound. The engineering is imaging a glowing specular seam and profiling its shape — before a model grades the bead.
CONTRIBUTION
What Faststream did.
The scope of the work, rather than a capability list.
Glare and glow control — filtered, timed capture that tames the weld's brightness and self-emission into a usable image.
Structured-light profiling — measuring bead height and shape, because weld geometry is three-dimensional and invisible to a flat image.
Seam tracking — following the seam on the robotic line so the inspection stays on the weld.
Geometry grading — bead height, undercut and fusion assessed against acceptance criteria.
Inline inference — a graded verdict produced at the pace the robot moves, logged per seam.
WHAT WAS HARD
The parts that consumed the schedule.
Rarely the subsystem that sounds difficult.
01
Imaging the weld at all
The specular, glowing seam is the crux; no model recovers a saturated, glared image, so most of the accuracy is won in taming the scene.
02
Seeing shape
Bead geometry and undercut are three-dimensional; capturing them needs profiling or structured light, which complicates the imaging further.
03
Trust for safety
Because the defects can be structural, the inspection has to be dependable rather than indicative, which raises the bar on both imaging and decision.
04
Keeping pace on the robot
Doing all of this while moving along the seam means the capture and decision have to keep up with the robotic motion.
OUTCOME
What resulted.
The seam, imaged
A usable image and profile of a scene that defeats a naive camera.
Shape as well as surface
Bead geometry and undercut graded, not just surface appearance.
Safety-relevant defects caught
Porosity and incomplete fusion detected dependably.
Inline and traceable
A graded verdict per seam, produced at line pace and logged.
Confidentiality
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture, confidential deliverables and commercial terms are not disclosed. Where a detail would identify a customer it is omitted rather than approximated. More is available under a non-disclosure agreement, within the limits the customer has agreed.
Because a weld seam is one of the more hostile scenes in machine vision. The metal is bright and highly reflective, so it glares, and a fresh weld can still be hot and glowing, which saturates the sensor. On top of that, the defects that matter — porosity, undercut, incomplete fusion — and the bead geometry are three-dimensional, so a flat image misses them. Taming that scene into a usable image and profile is the first and hardest part of the job, before any defect model runs.
02
Why is structured light needed?
Because much of what determines a sound weld is shape, not appearance. Bead height, undercut and the profile of the fusion are geometric, and a standard two-dimensional image cannot measure them reliably. Structured light or profiling reconstructs the seam's three-dimensional shape, letting the system grade the geometry against acceptance criteria rather than guessing structural soundness from a flat picture of a glaring surface.
03
Can't a good AI model handle a difficult image?
No model recovers information that the imaging destroyed. If glare washes out the surface or the weld's glow saturates the sensor, the defect simply is not present in the data for a model to learn or detect. That is why the work is imaging-first: filtered, timed capture and structured-light profiling produce an image and a profile a model can actually judge, and the model is only as good as that input.