Background
Quality control (QC) is essential in any manufacturing operation, but for the pharmaceutical industry it is mission-critical to ensure patient safety. Pharmaceutical companies have dealt with this challenge for many years by using human operators to visually inspect and ensure that the quality meets standards. However, it is a costly and ineffective endeavor.
Pharmaceutical manufacturers expect quality control processes to account for more than 20 percent of total production costs. They often operate at less than 15 percent of capacity and accept 5–10 percent of their production will need to be scrapped or reworked.† Human inspection is prone to errors in production lines that pack two pills per second. This process is not scalable without an associated increase in labor.
With the arrival of computer vision techniques powered by AI and deep learning, visual inspection can be digitized and automated to help improve the quality of the pills and lower the cost of operations.
Solution
Improve Automated Inspection with Computer Vision
In partnership with Accenture*, Intel developed an AI reference kit to help pharmaceutical companies automate the visual inspection of pills. Each kit includes:
- Training data
- An open source, trained, visual detection deep learning model
- Libraries
- User guides
- oneAPI components
The challenge with computer vision techniques is that they often require heavy graphics compute power during training and frequent retraining as new products are introduced. Optimizing on Intel® oneAPI toolkits helped improve performance and lower operational costs for training and inferencing. Using computer vision and SqueezeNet, the AI Visual QC model uses hyperparameter tuning and optimization to detect pill defects with 95 percent accuracy.
The AI Visual QC model was trained with the MVTec anomaly detection dataset (MVTec AD) and additional data derived from augmentation techniques, such as histogram equalization, letter box, random perspective, flip, scaling, and rotation. SqueezeNet classification architecture determines if the pill is good or defective with an associated confidence level of 0–100. If the confidence level is below your quality threshold, you can automatically route that pill for human inspection.
Technology
Optimized with Intel oneAPI for Better Performance
The AI Visual QC model was trained using PyTorch* and optimized using Intel® Optimization for PyTorch* for better performance. Reuse your PyTorch framework for model development without any code changes for training. The Intel® Distribution of OpenVINO™ toolkit was used to optimize inferencing for computer vision workloads across XPU and FPGA platforms.
Intel® oneAPI AI Analytics Toolkit was chosen to build training and inference models across a heterogeneous XPU architecture.
Components of the toolkit were used to test on Microsoft Azure* Standard_D4_V5 using 3rd generation Intel® Xeon® processors to optimize the solution.
Benefits
For data scientists, the AI Visual QC model means better model performance without sacrificing detection quality. With faster model training on Intel oneAPI toolkits, more machine learning models can be built and trained with less compute and lower costs. With Intel oneAPI toolkits, little to no code change is required to attain the performance boost. A 55 percent faster inference time means more accurate detection at scale when pills are moving at high speed on the production line.
For pharmaceutical companies, the AI Visual QC model means providing patient safety and meeting product quality and regulatory compliance requirements at scale. Production lines can now operate at higher capacity without constraints of quality control, which translates to higher profitability.
