Dashark | Dreamstime.com
Dreamstime M 80800241
Dreamstime M 80800241
Dreamstime M 80800241
Dreamstime M 80800241
Dreamstime M 80800241

Making Machine Vision Smarter With AI

April 13, 2020
Q&A with Landing AI's VP of Transformation Dongyan Wang takes a closer look at the challenges and keys to successfully leveraging AI as part of machine vision applications.

It is always exciting to watch emerging technologies evolve into the new normal. The key to this evolution is often rests with the ability to integrate the new offering as part of a proven technology. The adoption of artificial intelligence into machine vision is a prime example.

Read on as Dongyan Wang, vice president of transformation at Landing AI shares his perspective with IndustryWeek.

IW: Are there specific machine vision applications where AI has the biggest impact?

Wang: Visual inspection is an important step for quality control in manufacturing. Yet, many companies still rely on error-prone manual inspection or inflexible rule-based machine vision to find small defects like scratches or dents. 

While traditional machine vision works well on objects with minimal variations, it struggles in situations that have variability or high complexity. Traditional machine-vision solutions such as Automated Optical Inspection (AOI) are based on hard-coded rules or a golden image, which is comparing product images with a standard image without defect. As a result, it is not uncommon to see a rule-based system periodically classify a good part as defective (false positive or overkill) or vice versa, when there is any acceptable variation on the product. 

AI-powered vision detection provides speed, accuracy and repeatability at a lower cost. It excels at tasks like complex cosmetic inspections and segmentations, classification of defects, assembly verification, part location and challenging optical-character recognition. For example, we have demonstrated that AI performs much better than the traditional rule-based machine vision in lowering false positives by as much as 95%. 

IW: What do you see as the keys to successfully leveraging AI as part of a machine vision application?

Wang: There are three keys to successfully leveraging AI as part of a machine vision application.

1. Developing an efficient AI data acquisition and management strategy. A well curated dataset for deep learning training is the foundation for any AI project, so manufacturers need to establish a scalable, ongoing process where they can collect and manage data efficiently. This involves defect image acquisition, defect definition and clarification, data labeling, data verification and synthesizing rare defect data. The higher quality and closer to the real operating conditions your data is, the larger the chances of your project’s success will be.

2. Establishing a robust process to deploy and manage AI in production. Many companies can successfully create a PoC AI model for visual inspection fairly quickly. But the model is typically less than 10% of the overall effort needed for the first deployment. Manufacturers need to have a host of other tools like model monitoring and environment change detection to ensure the successful execution beyonce a PoC. Moreover, it is to the customer's best interest to continue collecting data, especially rare defects, and continuously update and improve the AI performance.

3. Bringing different stakeholders onboard to ensure your organization is ready to embrace AI. Implementing AI involves not only the technology but also organizational readiness. For example, an AI visual inspection system that helps detect defects affects many stakeholders from front line workers to process engineers. To keep projects on track, people must be brought onboard, and their workflow must be adjusted to take advantage of the AI. 

IW: What challenges do manufacturers need to overcome?

Wang: Building AI visual-inspection models has proven challenging and has held back many projects. The top challenges include:  

Small data. Whereas consumer internet companies may have big data from 1B or more users to train an AI system, a manufacturing plant may have only 10 or 100 or even fewer pictures of a particular defect they wish to detect. AI models developed for big-data problems do not work in small-data settings. 

Ambiguous-defect requirements. Identifying a defect can be subjective, and it’s common for two inspectors to disagree on qualifications. One inspector may consider a scratch to be problematic, while another thinks the same scratch is small enough to be ignored. These inconsistencies lead to mislabeled data, which can have a huge impact on the model's ability to accurately make predictions.

Changing environments and requirements. When changes in the external or internal environment occur, such as lighting and material, performance of AI systems degrades. Manufacturers also regularly change requirements that alter product appearance. For example, a factory may previously have considered 1 mm scratches acceptable, but new requirements now only accept scratches under 0.8 mm. 

Compounding complexities. No AI projects can achieve significant business values unless they can be deployed at scale. However, the need to scale introduces the problem of compounding complexities - every AI project is complex in and of itself, but when trying to scale a project to many stations, lines, and factories, the complexities only compound on themselves.

For example, a manufacturing company may decide to implement AI-powered visual inspection solutions across 100 factories to inspect hundreds of products, and each could have 100 potential defects. This can easily result in thousands of unique AI software models. This level of complexity is far greater than what any person can track. 

IW: How do you see the use and effectiveness of AI continuing to evolve within the machine vision space?

Wang: AI-powered vision application is still in the early stage of adoption in manufacturing, and many projects are stuck in pilots because of challenges ranging from a lack of data to knowing how to manage complex machine learning workflows. Going forward, we will see the emergence of verticalized, end-to-end platforms that enable users to move their projects out of the pilot and all the way to the finish line.

Over the next few years, we expect many companies will be able to deploy AI powered visual inspection much more broadly, with significantly less AI resources and short project life cycle, enabled by the vertical AI visual inspection platform and out of box solutions. 

Popular Sponsored Recommendations

Empowering the Modern Workforce: The Power of Connected Worker Technologies

March 1, 2024
Explore real-world strategies to boost worker safety, collaboration, training, and productivity in manufacturing. Emphasizing Industry 4.0, we'll discuss digitalization and automation...

3 Best Practices to Create a Product-Centric Competitive Advantage with PRO.FILE PLM

Jan. 25, 2024
Gain insight on best practices and strategies you need to accelerate engineering change management and reduce time to market. Register now for your opportunity to accelerate your...

Transformative Capabilities for XaaS Models in Manufacturing

Feb. 14, 2024
The manufacturing sector is undergoing a pivotal shift toward "servitization," or enhancing product offerings with services and embracing a subscription model. This transition...

Shifting Your Business from Products to Service-Based Business Models: Generating Predictable Revenues

Oct. 27, 2023
Executive summary on a recent IndustryWeek-hosted webinar sponsored by SAP

Voice your opinion!

To join the conversation, and become an exclusive member of IndustryWeek, create an account today!