Featured Product
This Week in Quality Digest Live
Innovation Features
Ian Wright
MIT and ETH Zurich engineers use computer vision to help adjust material deposition rates in real time
NIST
Having more pixels could advance everything from biomedical imaging to astronomical observations
Chris Caldwell
Significant breakthroughs are required, but fully automated facilities are in the future
Leah Chan Grinvald
Independent repair shops are fighting for access to vehicles’ increasingly sophisticated data
Adam Zewe
How do these systems differ from other AI?

More Features

Innovation News
Exploring how a high-altitude electromagnetic pulse works
High-capacity solution using TSMC’s 3DFabric technologies
EcoBell paints plastic parts with minimal material consumption
Study of intelligent noise reduction in pediatric study
Easy to use, automated measurement collection
A tool to help detect sinister email
Funding will scale Aigen’s robotic fleet, launching on farms in spring 2024
High-end microscope camera for life science and industrial applications

More News

Innovation

How AI-Powered Machine Vision Can Improve Your Supply Chain

Deep learning offers breakthroughs for quality control

Published: Tuesday, September 19, 2017 - 11:01

It seems to happen to every company, big or small, newcomer or seasoned expert. You ship a product design off to a manufacturer, and something goes wrong on the manufacturer’s side. The problem crops up in the design, production, or packaging, and leads to a bad apple in a batch of otherwise great products, or worse, a batch of unshippable product.

One of our friends, who oversees quality and product at an innovative eyewear startup in San Francisco, posed this very problem to us at Valkri Intelligence, where we develop deep-learning solutions for better quality control. We broke the problem down into three main points:
• How can we eliminate delays due to flawed optical products and guarantee near-perfect product quality?
• How can we provide insight into the quality control process?
• How can we achieve this through technology?

Let’s look at some of the ways you can integrate technology into your quality control process to address similar issues and achieve your quality, supply chain, and product goals.

Why might delays slip into your supply chain?

Let’s say you’re a manufacturer that, like our friend, works with an optics supplier. A huge part of the supplier’s job is quality control. They’ll do their best to deliver flawless lenses for your products by employing teams of quality control workers whose sole job is to manually detect flaws in the lenses by inspecting lenses through a backlight.

Unfortunately, this process is rife with human error. Different lenses have different thicknesses and characteristics, and requiring a human to adapt to these constantly changing characteristics is a sure way to introduce mistakes.

A single flawed lens that makes its way into an otherwise perfect batch will lead to customer dissatisfaction and decreased brand reputation. Human analysis is also slow and creates bottlenecks, driving up cost. So how can we improve the quality control process within these optics factories?

Our answer is to allow humans to work with a camera system with built-in intelligence. Such a system can perform the most critical aspects of quality control, while humans oversee the entire process. These machine vision systems have superhuman perception and can run 24 hours a day, without tiring, and autonomously adjust to changing lens properties.

Furthermore, instead of needing humans to tilt and inspect a lens at various angles and illuminations, a machine vision system can inspect an entire lens immediately and quickly detect flaws or scratches.

An intelligent machine vision system in this scenario would consist of industrial cameras that can take high-quality photographs of lenses in a production environment. These photos would then be processed by fast computer-vision algorithms in a computer that is connected to the camera. A human would be responsible for operating the system and staging lenses for analysis. Or the system could sit in an automated assembly line for full automation.

What is traditional computer vision?

Let’s back up a step and look at the technology behind machine visions systems. Computer vision (CV) is the study and development of algorithms that can give computers a higher level understanding of images and their visual environment. During the 1960s, researchers initially believed that CV would be easily solved. It has been almost 60 years since CV was first developed, and we still have a long way to go before computers can truly appreciate, interpret, and interact with the world like we do.

However, CV algorithms have made it possible for computers to perform certain tasks extremely well and with high precision. Incredibly accurate and fast solutions exist for facial recognition, people tracking, robotic navigation, and object recognition.

How does CV work, exactly? It’s an incredibly deep and varied field, so let’s concentrate on image processing and traditional computer vision techniques. In a subsequent article, we’ll look at cutting-edge vision methods that incorporate deep learning.

Before deep learning took center stage, CV often involved engineers and researchers devising custom algorithms for each task; that is, they custom-defined what to look for in an image. Although machine learning could be used to tell the difference between objects in an image, a human would still need to write an algorithm to extract useful information in a process called “feature extraction.”

For example, detecting people in a video might involve pixel-level operations like background subtraction to first isolate moving objects from video frames. Shape recognition, shape descriptors, and techniques like the histogram of oriented gradients could be used to detect the presence of people-like objects in the foreground. This entire process is defined by a human. Afterward, each people-like object could be classified by a machine-learning algorithm such as a support vector machine or a random forest to identify which objects are people and which are not.

How can these technologies help manufacturers?

In low-mix, high-volume manufacturing, computer vision can rapidly detect the presence of flaws, such as the presence of scratches or cracks in metals or plastics when viewed under proper illumination. It does this through a variety of methods, including:
• Template-matching algorithms, which compare a sample image to a template image to look for differences.
• Line detection, which identifies the presence of line segments.
• Simple connected-component analysis to determine when something shouldn’t be in the image.

Insight and analytics from autonomous quality control can also be of enormous benefit in machine calibration, design of experiments, or predictive maintenance.

What’s the catch?

The problem is that these algorithms only work well on a single task. If you wanted to perform quality control on any other range of products, or had a high-mix environment, you would need to write a completely different algorithm for each product.

Is it possible to create a computer vision system that could learn like a human? Could you, for example, present it with images that didn’t require feature-extraction algorithms, and “train” it to discover patterns on its own about what traits lead to a defective product?

Fortunately, such algorithms for CV exist, and this class of cutting-edge technologies is called deep learning. At Valkri, we found that traditional computer vision techniques didn’t give us the results we needed to help our friend in the optics industry, due to the wide variability of defects in the manufacturing process. Our answer to this challenge is deep learning, and we are developing intelligent solutions for lens quality control.

In our next article, we’ll talk more about deep learning, and how companies can use the technology to improve their manufacturing processes.

Discuss

About The Authors

William Hang’s picture

William Hang

William Hang, a co-founder of Valkri Intelligence, is a researcher in the Stanford AI Lab and previously worked on a machine learning project at Salesforce.

Zihua Liu’s picture

Zihua Liu

Zihua Liu, a co-founder of Valkri Intelligence, is a researcher in the Stanford AI Lab and ZhenFund Fellow, and a former researcher in the Stanford MobiSocial Lab.

Kevin Yang’s picture

Kevin Yang

Kevin Yang, a co-founder of Valkri Intelligence, is an AI engineer and designer who has worked with various Bay Area startups. He is also a ZhenFund Fellow.