What should you check when calibrating an egg classifier?

What should you check in the setup before calibrating?

The setup must be stable before any limit is touched. Diffuse, glare-free light matters more than any parameter, because strong glare creates false dirty eggs. If the light still changes during a shift, fix that first, as the guide on lighting for egg inspection explains.

1. Confirm that the camera is fixed above the conveyor and that its position does not change between calibration and production.
2. Confirm that the light is diffuse and that there is no strong glare on the shells.
3. Gather eggs photographed with the same camera, the same light and the same conveyor as production.
4. Confirm that the eggs appear whole in the frame, because eggs cut by the edge are marked unverified.

How do you label the calibration eggs?

A trained person labels each egg as clean, dirty, cracked or unverified, without looking at the system's decision. The labels define what the classifier should get right, so treat them with care.

1. Label each egg as one of Vigiovo's four classes: clean, dirty, cracked or unverified.
2. Decide disagreements together, and note the reason for the decision.
3. Deliberately include the doubtful eggs, such as light stains, fine cracks and eggs touching the edge of the image.
4. Do not label as cracked an egg whose crack only shows under candling, because the camera cannot see it and the label should follow the image.

Which limits should you tune, and in what order?

Vigiovo decides with fixed limits, and calibration tunes some of them. Change one at a time, in the order below, and run the same set after each change to see its effect.

1. Check the dirt limit: an egg is dirty when stains cover 3% or more of the visible shell, measured against the egg's own shading.
2. Check the crack limit: an egg is cracked when the total crack length reaches 0.8 of the reference radius.
3. Adjust the stain colour difference if light stains go unnoticed, or if the shadow of the shell is read as a stain.
4. Adjust the hue and saturation limits of the shell only if the colour strays from the default brown or cream, as happens with white or bluish shells.

The limits are relative to the reference radius, so they hold at any resolution. What changes from one line to another is the light, the shell and the camera, and that is what calibration captures.

How do you test the calibration before using it on the line?

Test the configuration on eggs that were not used for tuning. If the same set serves for both tuning and testing, the result looks better than it will on the line. Because classification is deterministic, the same image with the same configuration always gives the same result, so the test can be repeated at any time.

1. Set aside the test set before you start, and do not use it to tune limits.
2. Count how many clean eggs are rejected. The target is zero, but confirm the number on your own line, because the published results come from synthetic scenes.
3. List the eggs that sit close to a limit, since they show whether the boundary is crowded.
4. Repeat the test whenever the light, the camera position or the shell colour changes.

What should you record at the end of calibration?

A calibration that is not recorded cannot be audited or repeated. Store the final configuration together with the following:

1. Record the date and the version of the configuration in use.
2. Record the list of eggs in the calibration set and in the test set.
3. Record the test result by class, and the eggs that sat close to a limit.
4. Record the conditions of the setup: camera, position, light and shell colour.

Frequently asked questions

How many labelled eggs do I need to calibrate?

A few hundred, with variety. Include clean, dirty, cracked and doubtful eggs, because the borderline cases show where each limit should sit. A set made only of clean eggs does not teach the classifier to recognise defects.

Can I calibrate and test on the same images?

No. Set aside a test set before you start and do not use it to tune limits. If the same set serves for both tuning and testing, the result looks better than it will on the line, because the limits were chosen for those exact eggs.

Do I need to recalibrate when I change the camera?

Yes, if the change alters the position, the framing, the light or the image. Those changes shift the measurements, and the old limits stop holding even when the configuration is untouched. Run the test set again after any such change.

Sources

1. Como o Vigiovo classifica ovos sem IA / How Vigiovo grades eggs without AI, Vigiovo