Reinventing Quality Control in Raw Material Processing

Using computer vision to track particle size distributions in real time — reducing delays, improving accuracy, and enabling faster decisions.

The limits of traditional monitoring

Our client processes large volumes of raw materials via complex conveyor systems. They relied on lab samples taken only a few times per week, leaving critical data gaps. A key area of concern was the distribution of particle sizes within the material flow. Even subtle changes over time could indicate underlying issues such as equipment degradation, process inefficiencies, or variations in raw material quality. These shifts often required timely intervention from operators; however, the existing sampling approach lacked the frequency and resolution to detect them early.

To gain better visibility and improve responsiveness, the client required a solution that could deliver continuous, high-precision insights into particle behavior in real time. Using data from a camera monitoring one of the conveyor belts, we developed an optical system that accurately tracks particle sizes and analyzes their distribution over time. We validated the system's precision by comparing its results with laboratory measurement samples to ensure its reliability.

Implementing real-time monitoring in raw material processing

The project's biggest challenge was developing a reliable object detector to identify individual particles, even in low-quality footage with suboptimal lighting. Visual noise and slight camera movements made accurate detection even more difficult.

We addressed this through the following steps:

1. Model selection and transfer learning

We began using a pre-trained convolutional neural network and applied transfer learning to tailor it to our specific needs. A small set of manually labeled stock images featuring similar objects was used to fine-tune the model for the first iteration on real data.

2. Model refinement with real-world data

After testing the initial model, we enhanced its accuracy by incorporating images from the client's production environment into the training dataset. This made the model more effective at recognizing particles under varying conditions and better adapted to changes in material appearance.

3. Geometric correction

We applied a coordinate transformation to correct distortions caused by the camera's wide-angle lens and perspective. This was based on a calibration image captured on-site and standard computer vision techniques, enabling the system to infer real-world dimensions from the footage.

4. Statistical tracking

Finally, we developed a statistical module to track particle size distributions over time. The system could correlate particle sizes more accurately across the entire sequence by analyzing multiple video frames.

The impact of our solution

Even the initial versions of the model demonstrated impressive accuracy, achieving results within 5% of the laboratory samples. This means that our measurements were on par with those from the laboratory, falling within the expected range of statistical uncertainty.

Our solution not only ensures accuracy but also optimizes the use of computational resources, enabling exceptionally high time resolution down to just 100 milliseconds. This efficiency allows us to gain deeper insights into the structure of individual batches of material.

We helped our client transition from manual measurements to a fully automated system, improving accuracy, reducing delays, and enabling faster, data-driven decision-making.