TB-8529 · REV A · Technical newsheet

Machine Vision & InspectionDevice profile

MSTVision Puts 1 GB/s Photometric Stereo Against the 'AI Fixes Everything' Myth

At VISION 2026, MSTVision's Michael Stelzl called the idea that AI compensates for poor image quality a misconception, and showed a 1 GB/s GPU photometric stereo platform to back it up.

By Sophie Lindqvist3 min read610 words

Features

  • MSTVision's GPU-based area-scan photometric stereo platform processes up to 1 GB/s, per vendor claim at VISION 2026 in Stuttgart.
  • MSTVision offers three illumination options: segmented ring light (diffuse surfaces), flat dome light (shiny surfaces), and dome light as a compromise.
  • Michael Stelzl, managing director, said AI cannot compensate for poor image quality: "Many people think AI can do anything. From our point of view, that's not correct."
  • Both rule-based and AI-driven inspection systems perform better with high-quality images, according to Stelzl.
Why Image Quality Still Matters in the Age of AI
Device photoWhy Image Quality Still Matters in the Age of AI — Netti_Nu_Nu / pixabay

A GPU-based area-scan photometric stereo platform that processes up to 1 GB/s took center stage at VISION 2026 in Stuttgart last month — and it arrived bundled with a blunt message: AI does not excuse poor image quality.

The message came from Michael Stelzl, managing director at MSTVision, who told Vision Systems Design that one of the biggest misconceptions in machine vision today is the belief that AI can compensate for bad images. In his view, it can't. Successful machine vision systems still depend heavily on high-quality image acquisition, and that holds whether the downstream analysis is rule-based or AI-driven.

"Many people think AI can do anything. From our point of view, that's not correct," Stelzl said.

Why the argument matters for practitioners

The claim cuts against a common procurement shortcut. If a deep-learning classifier can be trained on whatever the camera happens to capture, why spend on premium optics and structured lighting? Stelzl's answer is that both classes of inspection systems — classical rule-based and AI-driven — perform better when the underlying images are good. Investment in imaging and illumination remains critical, and AI is not a substitute for proper imaging.

For an integrator, that changes where the budget goes. Money saved on a cheaper lens or a generic light source tends to reappear as labeling effort, retraining cycles, and borderline false accepts at the classifier stage.

The takeaway: the image chain is still the foundation, and AI is a consumer of image quality, not a generator of it.

What the platform actually does

Photometric stereo (an imaging technique that recovers surface detail by illuminating a target from multiple directions and comparing how the brightness changes) excels at revealing subtle surface structure — scratches, embossing, texture — that a single fixed light washes out. MSTVision's implementation is GPU-based and area-scan, and Stelzl quoted a processing throughput of up to 1 GB/s. That figure is a vendor-stated claim from the show floor; no independent test conditions were published alongside it.

The speed matters because photometric stereo traditionally trades frame rate for surface information. A 1 GB/s pipeline pushes the technique toward inline inspection, where parts move continuously and the system can't afford to pause for sequential exposures and offline fusion.

The takeaway: this is photometric stereo aimed at production-line speeds, not just lab surface analysis.

Three lights, three surface regimes

Illumination geometry decides whether a surface reads as informative or as glare, and MSTVision matches its platform with three illumination options, each tailored to a different surface class:

  • Segmented ring light — for diffuse surfaces, where directional segments can build up contrast from multiple azimuths.
  • Flat dome light — for highly reflective or shiny surfaces, where a broad, near-uniform field suppresses specular hot spots.
  • Dome light — positioned as a compromise between the two, covering surfaces that sit between matte and mirror-like.

Stelzl said the combination of high-speed processing and this flexible illumination range lets MSTVision optimize system performance across a wide span of inspection applications.

The takeaway: surface finish drives light choice, and having all three geometries in one toolbox widens the addressable part mix.

The question the platform raises

For teams already running AI-based inspection, the argument lands as a make-good decision: do you fix marginal image quality upstream — optics, lighting, photometric stereo acquisition — or keep absorbing it downstream in training and validation cost? MSTVision's position is unambiguous: fix it upstream. Whether that position converts into adoption will show in how many AI inspection projects specify photometric-grade acquisition from the start, rather than retrofitting it after classification accuracy stalls.

via Vision Systems Design (Source)

Filed under

  • photometric-stereo
  • machine-vision
  • ai-inspection
  • image-quality
  • illumination
Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering regulation, policy and their impact on Instruments, test and measurement at Testbench Report.

110 articles

Application notes

« Previous articleNext article »