Machine vision can catch manufacturing defects that may escape the notice of human inspectors, but traditional versions of these systems often have limitations.
That was the struggle at a leading global auto glass manufacturer, where water droplets and chips in the glass looked similar enough that the machine vision system couldn’t always tell the difference. Nearly one-quarter of the “flaws” flagged by the system turned out to be false.
While cameras and other hardware used in machine vision inspection systems keep advancing, it was the algorithms used in these systems that contributed to these false reject rates. When identifying “fuzzy” defects, often traditional vision algorithms have difficulty in differentiating between true defects and “pseudo-defects". However, by incorporating AI and deep learning capabilities, Mariner’s Spyglass Visual Inspection (SVI) can be trained by human quality experts to improve detection accuracy and overcome the binary nature of traditional defect detection algorithms.
How AI and deep learning create better results
The basics of machine vision systems are well established. With the right combination of cameras, lighting, and software, such a system can repeatedly inspect for a wide variety of manufacturing anomalies. Depending on the particular process of the manufacturer, products with detected flaws are often removed from the production line for further inspection or disposal.
Traditional defect detection systems use combinations of predefined, binary threshold values to identify defects. For example, if a spot is detected on a product that should not have any spots, the product is judged to be flawed. This type of binary detection, however, often doesn’t account for other factors that might produce a false finding, such as a spot being potentially nothing more than a piece of lint.
By using AI and deep learning technology, SVI uses image models that train it to tell the difference between actual defects and harmless anomalies. That’s why it can differentiate between glass chips and water droplets. To cite another example, in the case of an automotive interiors manufacturer, the solution can detect the subtle difference between stains and lint on automotive interior fabrics.
In both of these particular use cases, the accuracy rate of defect detection rose from 77 percent to 96 percent, reducing the false positive rate from “pseudo-defects” from 23 percent to less than 1 percent.
For inspections of products like these that are complex or include materials that may contain variations, markedly improved accuracy and reliability reduces the need for human follow-up inspections. By cutting secondary inspections and speeding up production lines, SVI can save considerable money for customers who use it.
Offering flexibility, customization, and scalability
While AI and deep learning are at the heart of SVI's inference engine, AI also offers flexibility in hardware requirements because it can utilize a variety of camera systems rather than requiring installation of a specific camera, which allows customer choice in cameras -- and also allows the system to use existing camera equipment in many cases. We also can work with customers who have aging camera systems or no existing cameras and recommend optimized hardware.
Multiple defect detection solutions on the market use pre-built, “universal” AI models that are good at some things but generally perform poorly on harder use cases. Mariner's team of on-staff data scientists, however, tailor each SVI model to each customer and product. This produces more accurate outcomes, especially for specialized manufacturing inspections. That means that although we cite here examples from glass and fabric manufacturing, SVI is easily implemented for different manufacturing scenarios and products due to the customizable AI and deep learning models.
Built with Microsoft and Intel technology
SVI utilizes Microsoft, Intel, and Nvidia technology to help deliver day-to-day reliability and speed for visual inspection. SVI uses the Microsoft Azure platform for centralized solution management, as well as security, scalability, and deployment ease. Additionally, the solution’s analytics are run on Azure, where they do not slow the real-time defect detection, which occurs on the Edge.
For training AI and deep learning models, we use the Azure cloud as well, and the Intel Deep Learning Boost (DL Boost) architecture improves these capabilities. Intel Xeon® processor-powered servers offer the performance needed for visual inspection without delays. Together, the joint Intel and Microsoft technology we use deliver edge-to-cloud services that also are secure and scalable for our customers.
Learn more about Spyglass Visual Inspection
Harnessing accurate machine vision technology can create more efficient manufacturing operations by reducing pseudo-defect detection and freeing up employees from needlessly re-inspecting products. SVI makes this achievable and can drive continuous improvement in quality by creating greater visibility into manufacturing processes. Learn more details about our Spyglass Visual Inspection solution and the technology that empowers it. Additionally, you can explore other solutions that incorporate Intel and Microsoft technology at TheIntelligentEdge.com.
This is a guest blog post provided to Mariner courtesy of Microsoft and Delightful Communications.