Comparison · Inspection technology
How do you keep an AI egg grader working over time?
Maintaining a grader means updating its decisions when the line changes: new lighting, a different shell colour, another kind of dirt. Tuning rules changes thresholds that the team can read, test and undo. Retraining a model needs new labelled examples, a new version and a new validation of everything it decides. For defects described by visible measurements, tuning is usually the simpler way to maintain.
Updated · 4 min read
What actually changes on the line over time?
Over time, the line changes: eggs arrive from another farm, shell colour varies with the season, the lighting ages and dirt appears in new forms. Any grader has to keep up with these changes, and the difference between the approaches lies in how that keeping up is done.
With rules, maintenance means reviewing thresholds and rules case by case. With neural networks, maintenance means collecting new examples, labelling them and training a new version of the model.
In both cases, the work does not end with the change. It is necessary to measure whether the result improved and whether something that was right before has become wrong.
Tuning rules or retraining: how do they compare in maintenance?
The table compares the maintenance work in each approach. It is a practical assessment, not the content of a standard, and the tuning column describes Vigiovo.
| Criterion | Tuning rules (thresholds) | Retraining an AI model |
|---|---|---|
| What changes | Thresholds and decision rules | Internal weights of the model, through new training runs |
| Who does it | The quality team, with calibration | Whoever trains the model, with new data |
| Data needed | Labelled eggs from your own line, to check the adjustment | Many labelled examples, including the new cases |
| How it is validated | Redo the calculation on labelled eggs and on cases close to the limit | Measure the new version on a separate set and compare it with the previous one |
| Effect on decisions already made | Only changes the decision of eggs whose measurement falls between the old and the new limit | A new version can change the decision for any image |
| Change record | Old and new limit, with date and reason | Model version, data used and validation result |
| Typical risk | A new kind of defect that no rule covers | A new data pattern that the model has not seen |
When is tuning rules the safer kind of maintenance?
Tuning rules is safer when defects can be described by visible measurements in the image, such as the dirt fraction or the length of a crack. Each change is small, can be tested against the same labelled eggs and can be undone. This holds for deterministic computer vision, where each limit has a known effect.
At Vigiovo, each decision records how far its measurement was from the limit. This shows where an adjustment would actually change something: if almost no eggs sit near the limit, moving it has little effect; if many do, it is time to review the calibration with labelled eggs.
The limitation appears over time. If a new defect has no rule, adjusting thresholds does not solve the problem. The team then has to write a new rule, which is engineering work and not calibration.
When does retraining a model still pay off?
Retraining pays off when the defect is hard to describe with rules, such as textures that vary widely and that no threshold separates well. In that situation the network learns the pattern from examples, and maintenance becomes the continuous collection of new cases.
The cost moves somewhere else. Instead of reviewing rules, the team labels examples, keeps the versions of the model and validates each new version against the previous one. Without that routine, a retrained model may improve one case and worsen another without anyone noticing.
The two paths can also be combined, with rules for the defects that are clearly described and a model for the rest. For the difference between the methods, see deterministic vision vs neural networks.
Frequently asked questions
Is tuning rules always cheaper than retraining?
- Not in every case. Tuning rules usually costs less when defects are described by visible measurements. When new defects appear with no rule, maintenance turns into writing rules, which is engineering work. The cost depends on how many kinds of defect the line presents.
When is a neural network still the best choice?
- When the defect is hard to describe with visible rules and there are enough labelled examples to train on. Maintenance then includes continuous collection of new cases and validation of each model version, and that has to be planned from the start.
Do I have to change the grader every time the lighting changes?
- Not always. A change in lighting calls for a new calibration with labelled eggs from the line, not necessarily a new rule. Repeat the calibration and check the cases close to the limit before returning to production.
Sources
How Vigiovo helps
Vigiovo grades eggs on the conveyor as clean, dirty, cracked or unverified, in real time, with deterministic computer vision: every decision comes with the measurements behind it, on an ordinary CPU.