Featured Product
This Week in Quality Digest Live
Metrology Features
Prashant Kapadia
Aid root cause analysis by detecting wear and tear in advance
Mike John
How metrology can bring the medical-device design and manufacturing process together
Steven Brown
21st-century standard candles at NIST
It’s actually the differences in a twin that are most useful

More Features

Metrology News
Meet the latest generation of LC xx6 encoders
On-demand pipe flow measurement, no process interruptions
Q-Span Gauging Systems improve efficiency, capacity, and real-time reporting of quality departments
Products range from software to scanners
See MobileCollect in Action in Booth No. 135413
Four new models available in TVM Series

More News

Michael Naber


Using AI to Enhance Machine Vision Inspection For the Auto Industry

Training data are a critical but difficult-to-achieve requirement for using AI

Published: Wednesday, February 9, 2022 - 13:03

The automotive industry is a critical part of the global economy, and the quality of the products it produces is essential for its success. Consumers demand high-quality vehicles that meet their needs, and companies in this industry must ensure that their products meet rigorous safety and performance standards. Quality control is therefore essential for automotive manufacturers, and automakers are constantly looking for faster, more accurate, and more human-free methods for inspection.

As it has in other industry applications, artificial intelligence (AI) is making inroads into automotive assembly and inspection. However, a major stumbling block is creating the training data required for AI inspection models. Defects are rare, and thus it’s hard to collect relevant imagery to train AI. It’s also expensive and time-consuming to annotate images. The automotive industry needs a better solution for collecting training data.

Today: Machine vision

Currently, automakers rely primarily on machine vision systems and human inspectors to make quality assessments. People and machine vision systems are used to inspect a wide range of automotive parts and components, including body panels, engines, and suspensions. However, both machine vision systems and human inspectors have their own sets of challenges.

Machine vision is a field of computer science and engineering that deals with the automatic interpretation of digital images. Machine vision systems are composed of several components, including cameras, lighting equipment, and lenses. Software analyzes these images to detect defects. In automotive manufacturing, machine vision systems are used to inspect products for defects and quality issues.

Machine vision systems can generally be tailored to detect specific problems. However, they’re expensive to deploy and require a lot of fine-tuning. The return on investment (ROI) from machine vision systems in automotive manufacturing is application-dependent. Some highly-automated processes, such as welding inspection, are better suited for the high-cost investment of machine vision systems than other applications, which are more artisanal.

In automotive manufacturing, it’s common to have a large workforce of human inspectors. However, training these employees can be an expensive and time-consuming process. In addition, the high turnover rate in this industry means that automotive manufacturers must continuously train new employees on how to inspect their products accurately.

Additionally, human inspectors may not be as accurate as machine vision systems, especially when they inspect automotive parts and components that are hard to reach or those with intricate designs. Human inspectors are also susceptible to fatigue and may not perform well when they work long hours.

Tomorrow: Leveraging machine vision with AI

Increasingly, automakers are opting to use their existing machine vision systems to test AI proof of concepts. By leveraging existing equipment, manufacturers can lower costs and thus increase ROI.

For example, Simerse can deploy deep-learning algorithms directly to existing machine vision cameras with either on-device processing or cloud-based inference. This means that your original investment in inspection hardware can be augmented with AI, even if the hardware wasn’t purpose-built for AI inspection capabilities. For example, machine vision systems used for tolerance inspection in vehicle-body manufacturing can also be repurposed to capture data for AI. Given the flexible nature of deep-learning algorithms, this system could even be redeployed with different lighting conditions to a different part of the factory.

How AI inspection is changing quality control

AI inspection is changing quality control in the automotive industry by increasing efficiency and accuracy. AI inspection systems are designed to be more cost effective than traditional machine vision or human inspection systems.

A picture containing text, clock  Description automatically generated
Annotated steel inclusion image. Source: Simerse

However, AI inspection has its own set of challenges. AI is not magic. It requires training data, which are images that are labeled for AI and machine learning.

Training data are what make AI inspection possible. The more training data an AI system has, the better it will be at detecting defects. Training data are made by annotating images of defects, i.e., drawing bounding boxes or polygons around the defect in each image. However, automotive manufacturers have a unique challenge when it comes to collecting training data.

The fact of the matter is that automotive defects are few. For example, defects in airbags are critical to find yet rare. This means that it’s difficult to collect relevant imagery for training AI models. In addition, annotating images can be expensive and time-consuming. Automotive manufacturers need a better way to collect training data for their AI inspection models.

Many automotive companies use their own custom-built datasets to train machine learning models for automotive inspection. These training data are used to create a model that detects automotive defects in real-time. However, these custom datasets have several limitations: They are expensive and time-consuming to build, the images they contain may not reflect production conditions due to calibration issues or lighting problems, and data annotation can cost up to $30 per image.

Alternatively, some automakers look to pretrained AI models or training datasets provided by companies like Simerse to overcome this hurdle. Simerse creates synthetic training data that can be used to train real-world AI algorithms. This approach can be much quicker and more cost-effective than real data collection by an automaker. However, some automakers prefer to take an in-house approach and collect these training data themselves. Open-source tools do exist if a manufacturer prefers this route.

What to consider when collecting training data

It’s important to remember that the automotive industry is global, and vehicle configurations change frequently. Parts found in one vehicle may not be found in another, and the specific part may change during vehicle model’s years in production. This means that the AI inspection systems, and thus their training data, must be able to handle a wide range of automotive products and components.

As a helpful guide, automotive manufacturers should consider the following when collecting training data for their AI inspection models:
• The type of defect that must be detected
• The location of the defect on the part or component
• The size and shape of the defect
• The color of the defect

This information must be recorded to teach the AI how to recognize that particular type of automotive defect. Because the in-house data collection and annotation process can be burdensome, many automakers opt to work with an outside vendor like Simerse.

Fundamentally, the best way to overcome the data bottleneck is to have a plan. We know that training data are rare and expensive to annotate. That means AI inspection efforts for your auto factory should be carefully targeted to the problems that yield the highest ROI. Consider using a proof of concept to validate the costs and savings of AI inspection for a particular type of defect before attempting a factorywide deployment.

It’s also important to remember that you’re not alone. Almost all major industries, from autonomous driving to robot vision, suffer from the training data problem. But with the right attitude, proof of concepts, and data-driven decision making, your AI investments can yield dividends now and in the future during quality inspection.

Why work with a vendor?

For many automakers and auto suppliers, the core business is making vehicles, not developing AI for quality inspection. Even the largest automakers can struggle to hire the talent necessary for AI quality control. Moreover, an entirely in-house approach often leads to reinventing the wheel. With AI technology moving at lightning speed, it’s beneficial to partner with a vendor whose entire business is focused on quality inspection.


AI inspection is changing quality control in the automotive industry by making manufacturers more efficient and accurate. AI inspection systems are designed to be cost-effective for automotive companies, but they still require training data that are specific to their automotive products and components.

Automotive defect-detection training data present their own challenges because automotive defects are rare, and relevant imagery is hard to collect. Automotive manufacturers must find a way to collect training data or work with a trusted vendor for their AI inspection models if they want to remain successful in their industry.


About The Author

Michael Naber’s picture

Michael Naber

Michael Naber is the founder and CEO of Simerse, an AI inspection and defect detection company.