How vision quality control on fruit and vegetables works
Vision quality control judges each product on the line with cameras and an AI model that has learned the standard from examples judged by quality experts. In the system of TOS Robotics, one image with one or more products is analysed in 0.02 seconds.
How it works
- Light and camera. Even light and a fixed camera position give images that can be compared. Many vision problems start here and not in the AI model.
- Find the product. The model separates each product from the belt and from its neighbours.
- Judge. For each product the model gives a class or a score, such as good, second class or reject, or a colour grade.
- Act. The line acts on the verdict. A reject is pushed out, or in a robot cell not packed.
- Store. The verdicts can be counted per batch, for a quality report per supplier.
What a camera can and cannot see
| Can see | Cannot see with a normal camera |
|---|---|
| Colour and colour spread | Firmness |
| Size and shape | Internal browning |
| Marks, cuts and pressure spots on the visible side | The side that lies on the belt |
| Visible decay and mould | Taste and sugar content |
| Stalk present or missing |
For the hidden side the product is turned or seen from several sides. The inside needs other sensors.
How the model learns your standard
Quality specifications can differ per retailer, so the model is trained on examples that your own quality experts have judged. They sort a set of products into classes, and the model learns to make the same split. When the standard changes, the examples change and the model is trained again. You do not write rules such as “reject when the spot is larger than five millimetres”. You show what you reject.
To start, choose a few classes and collect examples of each through the season, with extra attention for rare defects. Let two experts judge the same set. Where they disagree, the standard is not yet clear and the model gets no consistent example to learn from. Measure the result on products the model has not seen.
At TOS Robotics
On the bell pepper line, TOS Robotics builds the check into the singulator, in front of the robots. Four cameras photograph each pepper from different sides while it falls, and one image is analysed in 0.02 seconds. The design lets a rejected pepper leave through a separate chute before it reaches the robot. The model is being trained on the peppers of that line and does not sort in production yet. Read more on vision quality control.
Questions and answers
Can vision quality control measure ripeness?
Only as far as ripeness shows on the outside, such as the colour. Firmness and internal quality need other sensors.
How many example products does a model need?
Enough to cover each class you want to recognise, including the rare defects. The rare defects usually decide how long collecting examples takes.