By Clément BERNARD, Vision engineer and trainer at CODA Systèmes
A false rejection in industrial vision can be diagnosed in 15 minutes by following 5 steps: determine if it is new or structural, check if it comes from the image or the algorithm (70% of cases come from optics or lighting, even before processing), relate the image to the actual variations in the process, understand precisely what the criterion measures before adjusting it, and then test the robustness of the threshold on borderline parts rather than just its immediate performance. Modifying a threshold without following this method masks the symptom without addressing the cause, and weakens the system in the long term.
When the false rejection rate skyrockets, the classic reaction is to adjust the thresholds, widen a tolerance, or bypass the camera.
👉 Bad idea.
❌ Why touching the thresholds too early worsens the situation
A threshold is only the digital translation of a human decision.
Modify a threshold without knowing:
- what the algorithm actually measures,
- how the image is constructed,
- what variations are normal in your process,
Returns to masking the real problem, weakening the system, and creating invisible drifts.
👉 A "tuned" system that is not understood becomes an unstable black box.
Here is a simple, practical method, in 15 minutes, to find out where the real problem is.
⏱️ Step 1: Is the false rejection new or structural? (2 min)
- Has it always existed?
- Did it appear after a change (part, pace, lighting, cleaning)?
👉 If it is old → design issue
👉 If it is recent → drift problem
🔍 Step 2: Does the rejection come from the image or the processing? (2 min)
One must ask oneself a single question:
In the raw image, do I, as a human, clearly see the flaw or the drift?
- ❌ No → optical problem / lighting / positioning
- ✅ Yes → algorithm problem / criteria
⚠️ 70% of false rejections occur before the algorithm.
✅ Step 3: Connect the image to the actual process (5 minutes)
A vision image never lives alone. One must ask oneself:
- What really varies on the line?
- Does the part always arrive in the same way?
- Are there any mechanical, thermal, or material tolerances?
👉 If the process is not controlled, no threshold will compensate for it sustainably.
⚠️ Step 4: Understand the criterion BEFORE adjusting it (3 minutes)
Before changing a threshold, require a clear answer to these questions :
- What does this criterion actually measure ?
- Why was this value chosen ?
- What is the acceptable margin?
If no one knows how to answer precisely:
👉 The problem is not the threshold, it's the method.
🧪 Step 5: Test robustness, not performance (3 minutes)
A good threshold is not one that only reaches 100% today. It is the one that :
- accepts the good "limit" parts,
- rejects the bad "clean",
- resists the realistic drifts of the terrain.
👉 Otherwise, it amounts to solving the problem for the present and forgetting the future.
It is therefore necessary to test with :
- 10 good "limit" pieces
- 10 clearly bad pieces
- 10 deliberately degraded images
Moreover, creating a small confusion matrix is ideal !
CONCLUSION
👉 A false rejection is not resolved. It is understood.
Touching the thresholds without understanding :
- the image,
- the process,
- the criterion,
amounts to putting a band-aid on a symptom.
🔎 And concretely ?
That's exactly what we do during our diagnostics :
- understand what the camera actually sees,
- link the image to the industrial process,
- identify whether the problem comes from the optics, the method, or the criterion,
before changing the slightest threshold.
At CODA Systèmes, we assist manufacturers in diagnosing their false rejections related to image, process, or criteria before any threshold modifications. ➡️Contact us for an audit of your vision system.
FAQ
A threshold is the digital translation of a human decision. Modifying it without understanding what the algorithm actually measures, how the image is constructed, and what variations are normal in the process amounts to masking the problem rather than solving it — and creates invisible drifts.
Just look at the raw image: if a human does not clearly see the defect or drift on it, the problem comes from the optics, the lighting, or the positioning — not from the algorithm. This is the case for 70% of false rejections.
By confronting it with three sets of images: good "borderline" pieces, clearly bad pieces, and intentionally degraded images. A confusion matrix on these three sets allows us to verify that the threshold withstands realistic field deviations, not just the current situation.
No. First, it is necessary to understand the picture, the process, and the measured criterion — changing a threshold without this understanding is like treating a symptom without identifying the cause.