Context and Challenges
Quality control in the tobacco industry is extremely rigorous, often favoring the rejection of potentially good products over the risk of passing defective ones. If the defect detection system lacks precision, unnecessary product waste can quickly translate into thousands of euros lost daily. This project specifically required evaluating the quality of individual tobacco items as well as entire packaged products, each presenting distinct challenges and necessitating different approaches.
We faced a few challenges along the way:
- High Throughput: Maintaining high inspection throughput with minimal latency was essential to ensure the system could be installed on the machine without lowering current production rate.
- Individual vs. Packaged Items: Distinct approaches were needed to inspect both individual items and entire packs, each scenario having its own set of complexities.
- Product Variability: High variability in product appearance necessitated flexible yet reliable inspection methodologies.
Solution
To address these challenges, we developed tailored inspection solutions specifically optimized for individual tobacco items and packaged products.
Individual item inspection:
- Leveraged existing camera and lighting setups.
- Adopted a hybrid solution combining semantic segmentation models with explicit rule-based logic to accurately assess product quality from segmentation outcomes.
- Quantized and optimized model to run superfast on the Hailo-8 accelerator, with negligible accuracy loss.
Packaged Product Inspection:
- Developed an optimal lighting setup designed to reveal even subtle defects while disregarding the unknown or variable packaging appearance.
- Applied a proprietary anomaly detection model to implicitly differentiate between acceptable and defective products without requiring explicit definitions.
- Packaged everything to run on a standalone NXP i.MX8M Plus module, affordable and reliable industrial grade hardware.
Results
The implemented deep-learning-based model effectively reduced unjustified product rejections by approximately 10%. It was successfully deployed on an industrial grade embedded platform equipped with neural network acceleration capabilities, processing multiple production lines simultaneously with an impressive performance of under 15 ms per line.
