ARNEB by Altronis

ARNEB · Visual inspection

Catch the defects, keep the good parts. Teach it with good parts only.

ARNEB marks the spot on any part that differs from your good ones, and your inspector makes the call. It checks by photo, and on its 3D station by depth and colour too. It runs on a PC at your site, so nothing leaves your factory.

  • 2D photosmeasured on a public dataset
  • 3D depth and colourin development on simulated parts, real camera next

30 minutes, no slides, no obligation.

Fig 0 · ARNEB's 3D check, replayed on simulated partsDrag to turn
543/600

defects caught on public VisA photos

15/510

good parts sent to a person, the rest pass

0

defect photos needed from you

1 PC

at your line, nothing sent to a cloud

2D photo check, on the 6 public VisA categories we tuned on, so a new part likely scores lower. 3D depth and colour checks: in development, no accuracy figures yet. How we measured

01 · How it works

It flags what differs from your good parts.

Many systems need defect photos, which are slow to collect. ARNEB starts from the good parts you already make.

STEP 1 · TEACH

Show it good parts

About 200 good parts under your camera and light, with a depth scan of each on the 3D station. No defect photos, no drawing boxes.

STEP 2 · CHECK

It checks every part

Anything that differs from your good parts gets a score and a marked spot.

STEP 3 · SORT

Pass, review or fail

Below the review level the part passes. In between, it goes to a person with the spot marked. Above the fail level it is marked as a fail and logged.

STEP 4 · DECIDE

Your inspector makes the call

The inspector sees the marked spot beside a lined-up good part and presses Good part or Defect. Every decision is logged.

02 · 3D checks

Some defects only show up in 3D.

A warp, a dent or a missing feature can look fine in a photo. The 3D station adds a depth scan and a colour check.

In development. Simulated metal parts and a simulated depth camera for now. A real camera is next, and there are no 3D accuracy figures until then.

Colour photo of a simulated metal bracket that looks like a good part

A · The photo

Looks like a good part.

Depth scan of the same bracket in shades of teal, darker where the surface is higher, with height lines

B · The depth scan

Darker is higher. The bottom of the face sits higher than the top: a warp.

The same bracket with the warped area shaded red by ARNEB and framed by corner marks

C · What ARNEB marked

The flat-face check marks the warp as a fail: score 13.3, fail level 3.44.

Fig 2.1 · One warped bracket, three viewsSimulated part and camera · real ARNEB engine
Depth AI

Shape that changed

Marks dents, missing features and anything at the wrong height.

Flat-face check

Faces that should be flat

Marks warps and steps on faces that should be flat.

Colour AI

Marks on the surface

Marks scratches and spots. It can also catch a dent or a missing feature by its shading, as with parts 03 and 05 in Fig 0.

Where it is weak so far

Small warps, dents and burrs are often missed, or sent to review instead of marked as a fail. We are working on these before the real camera test.

03a · Evidence · 2D photos

In 2D, every flag comes with a good part to compare.

ARNEB's own outputs on test photos it never learned from, in categories we tuned its settings on. They come from VisA, a public Amazon set of circuit boards, capsules, candles and foods like cashews. It is used to compare inspection AI, so anyone can verify the results. Move over a photo to line up the same spot on both.

Fig 3.1 · The 200 good photos ARNEB was taught on for one circuit board (VisA pcb1)Real dataset photos
200

photos of good boards: the whole teaching set for this part.

0

photos of defects. It has never seen a broken board, and it still flags one.

FLAGGED · part on the line Circuit board on the line with a component missing, marked by ARNEB
GOOD PART · lined up to compare Matching good circuit board with the same spot marked
REVIEW · sent to a person
A component is missing

The good board has a part in this spot. This one does not. ARNEB scored it above the review level and sent it to a person with the spot marked.

Score against this part's levels

Score 3.58. Review level 2.64, fail level 4.01.

Photos: VisA dataset (Amazon), CC BY 4.0. Zou et al., ECCV 2022.

03b · Evidence · the numbers

Defects caught, good parts kept moving.

2D photo results on the 6 public VisA categories we tuned ARNEB on. What counts as caught was written down before the run.

Defects caught

of 600 defect photos · higher is better

Before tuning 436 · 72.7%
Tuned by Altronis 543 · 90.5%
0300600

Good parts sent to a person

of 510 good photos · lower is better

Before tuning 19 · 3.7%
Tuned by Altronis 15 · 2.9%
02551 (10% of 510)
By category · VisA tuning set · after tuning
CategoryDefects caughtGood parts sent to a person
Candle89 / 1006 / 100
Capsules97 / 1000 / 60
Cashew83 / 1000 / 50
Macaroni (1)94 / 1002 / 100
Circuit board (pcb1)94 / 1005 / 100
Circuit board (pcb2)86 / 1002 / 100
All 6543 / 60015 / 510

How we measured

  • 200 good photos per category to teach. No defect photos.
  • Pass, review and fail levels set from good photos only.
  • "Sent to a person" counts every good photo that landed in review or fail.
  • Written down before the run: what counts as caught.
  • "Before tuning" is the ARNEB build of 17 Nov 2026, run on the same photos.

What these numbers are not

  • These are tuning categories. We adjusted ARNEB on them, so results on a new part are likely lower.
  • Next: 6 categories it has never seen, kept untouched for that test.
  • No speed quoted yet. It will be measured on the station PC in that test.
  • Your part is not VisA. Results depend on the part, the camera and the light.

04 · Fit

Where ARNEB fits, and where it doesn't.

A good fit

  • Surface defects you can see in a photo: missing parts, chips, cracks, marks, stains, wrong shape
  • Parts that look much the same each time, under a fixed camera and steady light
  • Lines where you have plenty of good parts but few photos of defects
  • Inspection where a person still signs off and you want them to look only where it matters

Not a fit today

  • Measuring tolerances the way a CMM does
  • Seeing inside a part
  • Height and flatness on a real line, until the 3D station is tested with a real depth camera
  • Safety-critical final sign-off with no person in the loop
  • Moving cameras or light that changes through the day

05 · Setup

A PC at your line. We help you choose the camera and light.

No calls to a cloud service. It needs the internet once, at setup.

Camera input

Folder, RTSP or USB

Watch a folder, read a network camera over RTSP, or take a USB camera, with triggers.

GenICam industrial cameras: in testing
Operator screens

English and Japanese

Teach, live, check parts, history and recipes. Each part number gets its own recipe.

Records

Every decision logged

Each decision goes in an audit log with the inspector's name and time. Retention never deletes parts waiting for review.

Access

Stays on your network

Photos and results stay on the PC. Anything that changes data needs an API key.

Integrations

REST API and CSV today

Pull results into your own systems through the API or CSV.

PLC and OPC UA: planned, not built
History

Find any past part

Filter by date, outcome, recipe, lot and serial number, and export to CSV.

Illustration of an inspection station: a camera on a stand over a part, a PC screen beside it
Fig 5.1 · An inspection station at the lineAI-generated illustration
Fig 5.2 · Check screen, a 2D camera part selected. Its score is real engine output on a VisA photo; serial and lot are sample values. 3D parts in the queue are simulatedScreen design · Tap to open full size

06 · FAQ

Questions inspection teams ask us.

Do I need photos of defects to set ARNEB up?

No. It is taught on about 200 photos of good parts and flags anything that differs, so it can catch a defect you have never photographed.

Does ARNEB reject parts on its own?

No. It marks and logs each part but does not move parts off the line: PLC connections are not built yet. Anything between the review and fail levels goes to your inspector with the spot marked.

How accurate is ARNEB?

On the 6 public VisA categories we tuned on, ARNEB caught 543 of 600 defects (90.5%) and sent 15 of 510 good parts to a person (2.9%). These are tuning categories, so a new part likely scores lower. Next is a test on 6 categories it has never seen. See the full table.

Does ARNEB check parts in 3D?

In development. The depth AI, flat-face check and colour AI run on simulated metal parts with a simulated depth camera. No 3D accuracy figures until a real camera is tested. See the 3D checks.

Do my photos leave the factory?

No. It runs on a PC at your site and makes no calls to a cloud service. It needs the internet once, at setup.

What hardware does ARNEB need?

A PC at the line, a fixed camera and steady light. We help you choose the camera and light for your part. No speed figure yet.

Which cameras does ARNEB work with?

A watched folder, an RTSP network camera or a USB camera, with triggers. GenICam industrial cameras are in testing.

Can ARNEB connect to our PLC or MES?

A REST API and CSV export today. PLC and OPC UA connections are planned, not built yet.

Is the operator screen available in Japanese?

Yes. The teach, live, check, history and recipe screens are in English and Japanese.

Who makes ARNEB?

Altronis Pte. Ltd., an AI company in Singapore. Write to hello@altronis.sg.

Tell us about the part you check today.

Describe the check, or bring a few photos of good and bad parts. We will tell you plainly whether ARNEB fits.

30 minutes, no slides, no obligation

Book 30 minutes

hello@altronis.sg