AI Camera System for detecting and counting plastic bottle caps on a conveyor belt, with an air-based defect sorting mechanism for plastic cap manufacturers.

How the brief came about

The brief in this case was for a camera to inspect and count plastic water-bottle caps on a conveyor, in place of manual re-inspection and counting. The customer is a plastic cap manufacturer. Caps leaving the plastic injection machine travel along a conveyor, and staff then re-inspect and count them by hand.

The defects the customer wants rejected are flash and incompletely moulded caps. The customer also wants counts of good and defective caps as production reporting data, not just the removal of defective caps.

Requirements and constraints

The system must inspect the caps passing along the conveyor and support more than one cap colour. Its main requirements are as follows:

• Inspect every cap passing along the conveyor and find caps with flash or incomplete moulding
• Reject defective caps from the conveyor as soon as they are detected
• Support both white and blue caps
• Count good (OK) and defective (NG) caps for production reporting
• View the live image and counts, and configure the system, in a web browser on a tablet or PC

How does this system inspect and count caps on the conveyor, and how is the rejection of defective caps designed?

The system uses a LearningCam camera (Inspection Cam model) above the conveyor. The camera captures images of the caps continuously, uses an AI model to judge whether each cap is good or defective, and then counts them. Under the design approach, when a defective cap is found, the camera signals the Controller and Reject set (control and rejection unit) to blow that cap off the conveyor with air.

1. 1

   Caps travel past the camera on the conveyor

   Caps leaving the injection machine travel along the conveyor, passing under a LearningCam camera (Inspection Cam model) held above the conveyor by a camera jig/fixture.

2. 2

   The AI model judges OK or NG

   The camera captures images continuously. An AI model trained on real images of the factory's caps checks whether each cap has flash or incomplete moulding, then judges it as good (OK) or defective (NG).

3. 3

   Defect Sorting Mechanism Using Air Blowing

   Under the design approach, when a defective cap is found, the camera signals the Controller and Reject set so that a solenoid valve (an electric valve that opens and closes the air supply) blows that cap out with air. Good caps continue as normal.

4. 4

   Counting and display

   The software built into the camera shows the live image with OK/NG counts. Users open a web browser on a tablet, screen or PC and enter the IP address (the camera's address on the network), connecting through a Wi-Fi 6 router and an IoT Gateway (a device that connects on-site equipment to the network).

5. 5

   View the inspection report

   The production team opens the inspection report page (Inference Report) to summarise results by time period and by model, and can start or stop inspection from the production office over the network.

We chose a camera with inspection and counting software built in, so users can view results in a web browser without setting up a computer beside the line. We trained the model on real images of the factory's caps because the defects to catch are those that arise from this factory's injection moulding.

Deliverables

We supplied the equipment, installed the system on the customer's conveyor and trained an AI model for white caps. The items in scope are as follows:

• 1 LearningCam camera (Inspection Cam model) with built-in software that shows the live image and OK/NG counts
• Controller and Reject set with a solenoid valve for rejecting defective caps with air
• Wi-Fi 6 router and IoT Gateway to connect the camera to a tablet or PC in the factory
• Installation of a camera jig/fixture above the conveyor and on-site system configuration
• AI model for inspecting cap defects, trained on real images of the factory's caps
• Inspection report page (Inference Report) for viewing results by time period and by model

Results confirmed by on-site testing

In on-site testing from January to March 2026, the LearningCam camera (Inspection Cam model) inspected and counted about 398 white caps per minute, continuously for 120 minutes, and met the criteria set in the internal test report form. This result is a test of inspection and counting. It is not yet an acceptance result for the complete rejection system.

Blue caps tested with the white-cap model have not yet reached the criteria, so there is a plan to train separate models by cap colour. Work is in progress to adjust the model for other cap colours and to confirm delivery of the complete system.

During testing, a weak Wi-Fi signal made the live image opened from the production office stutter, while start and stop commands and the counts remained usable.

What we learned

This work offers the following principles for AI camera inspection of plastic injection-moulded parts:

• A model only works with the kind of part it was trained on. In this work, the model trained on white caps has not yet reached the criteria when used on blue caps, so a new cap colour or variant must be treated as a new round of model training and testing.
• Confirm inspection results separately from rejection results. Cap inspection and counting can meet the criteria first, but the air rejection set needs its own acceptance test, because the result depends on both the model's decision and the air-blast mechanism.
• Core work should not depend on the live image. In this work, a weak Wi-Fi signal made the live image stutter, but start and stop commands and the counts remained usable, so the signal at the viewing location should be surveyed before installation.

How can you apply this approach to your plastic injection line?

If your line produces caps or small plastic injection-moulded parts that travel on a conveyor, you can use the same approach as this case: a camera above the conveyor, a model trained on real parts, OK/NG counts and the design approach for an air-blast rejection set. What needs adapting is the model for each part colour or variant, and the mounting positions of the camera and the air nozzle.

• Prepare good parts and real defective parts of each colour or variant for the team to photograph for model training
• Agree the acceptance criteria before testing, including conveyor speed, test duration and measurement method, looking separately at the rate of detecting defective caps and the rate of wrongly rejecting good caps, not at overall accuracy alone
• Start with the colour or variant produced in the largest volume, and add models for other colours once it meets the criteria
• Test under the real lighting on site, and mount the camera so that the image angle stays fixed