When a leading medical device manufacturer sought to leverage deep learning in its inspection of transparent medical components, EZ Automation offered a quickly deployed, highly accurate automated vision inspection system.
The Problem
Medical device manufacturers have long relied on machine vision for quality assurance. But traditional rules-based vision systems do not adapt well to environmental changes — such as shifts in lighting — or to variability in inspected items.
One leading contact lens manufacturer discovered these limitations when it sought to automate the inspection of its transparent products and packaging. Its existing rules-based vision systems required unique algorithms for each possible defect type. Although the manufacturer committed significant time and expertise to the task, the rules-based system was limited in its ability to detect novel defects.
Transitioning to deep learning–based inspections offered a more robust solution that could identify previously undescribed defects based on learned patterns rather than predefined rules. However, this development process also presented obstacles. Training deep learning models requires manual labeling of thousands of images illustrating defective and acceptable conditions. The challenge was compounded by the need to identify rare but critical defects that occur in less than 1% of inspected items.
With production volumes exceeding five billion units annually, the manufacturer needed a solution that could simultaneously handle multiple defect categories while reducing the time and resources required to develop and deploy new inspection models.
The Solution
The manufacturer deployed EZ Automation’s EZ Eye platform, a deep learning–based inspection system specifically designed to address the challenges of complex quality assurance applications in regulated industries.
Underlying EZ Eye’s unique capability is its intelligent data curation methodology, which significantly accelerates the training process for deep learning vision systems. Rather than requiring human operators to manually sort through massive image datasets, EZ Eye applies algorithms that identify patterns in grading behavior and automatically categorize similar images. This allows quality engineers to focus their attention on novel or ambiguous cases that genuinely require expert judgment.
The platform proved particularly effective for detecting even small amounts of foreign matter in the medical device components. By intelligently prioritizing which images required human review, EZ Eye maintained model accuracy above 98% while dramatically reducing the human effort required during the training phase.
The Benefits
EZ Eye’s technology-agnostic architecture enabled seamless integration with the manufacturer’s existing process control systems. The platform can accommodate client-defined tolerances and quality thresholds, which allowed the manufacturer to tailor inspection criteria to performance and regulatory requirements.
Depending on the complexity of the defect categories, the medical device manufacturer achieved a 55% to 97% reduction in the time required to train its deep learning models. The EZ Eye platform’s ability to handle more than 10 distinct defect classes within its deep learning framework further eliminated the need to develop and maintain separate rules-based algorithms for each defect type.
Even with dramatic reductions in development and training time, the system’s detection accuracy exceeded 98% across multiple defect classes, including foreign matter, dimensional variations, and packaging defects.
Following successful proof-of-concept testing, the manufacturer expanded the EZ Eye platform across dozens of production lines at multiple manufacturing sites, demonstrating the solution’s ability to scale from laboratory validation to full production deployment.