We started our collaboration by visiting the customer’s plant to study their current manufacturing and coating inspection processes, IT infrastructure, and defect validation requirements and guidelines.

Based on our findings, we conceptualized a computer vision solution powered by neural networks. Aiming for non-disruptive production pipeline adjustments, we suggested enabling AI detection at the final production process stage with the help of a high-magnification camera that would capture container images and send them for analysis to the AI solution. Depending on the results, the container would be either approved for shipment or deemed defective.

One of the key project requirements was ensuring that the AI model could segment coating areas based on coating quality. To enable this, we worked with the customer to prepare a dataset containing about 1,000 labeled images to properly train the AI model. Moreover, we established a pipeline for automatic model retraining so that it learns based on live data over time, adapting its outcomes without human intervention.

We divided the solution deployment process into several stages. Initially, we launched the AI solution across all production lines to run in parallel with the existing defect detection workflow so we could observe its operation. We analyzed and validated the system’s output for each image, such as whether a defect was detected, whether there were no defects, and where the defect was found, and retrained the solution based on identified edge cases to have it meet the expected accuracy threshold.

After going through three review and retrain cycles, we managed to achieve 99.3% system accuracy, which we verified using 10,000 products. Finally, we received approval from the quality control department to put the system into full operation,replacing the previous process.

The delivered AI solution automatically performs the following defect detection tasks:

• Estimates the area of the protective coating and its share in the total container area
• Identifies coating gaps and their percentage ratios
• Measures coating thickness and detects thin layers
• Makes the decision whether the product is defective or not
• Sends automated alerts to the conveyor system to remove defective containers

We rolled out the AI solution to the customer’s on-premises environment to ensure it works without interruption regardless of the internet connection and enables real-time defect detection. Post-deployment, our team monitored the solution to prevent model drift and collaborated with the customer’s production and IT teams to develop new features, enhance the solution’s detection capabilities, and provide ongoing advisory support.