Simple rule-based algorithms have served the visual inspection market for many years, but as their limitations become apparent, the need for more sophisticated software increases. In this article, Miron Shtiglitz, head of product management at visual inspection company Lean AI, discusses his two eras of machine learning in visual inspection, currently featuring various levels of semi-supervised deep learning. It claims that we are entering a new third era. .
rule-based system
Many of the most popular solutions available in the visual inspection market can be characterized as rule-based. They work with different algorithms and the parameters are set by experts. A simple example might be a rule that counts the number of black pixels in an image and flags the image as defective if it exceeds a certain number.
Such solutions have been in use for at least 25 years and have proven to be highly effective for simple tasks. However, for more complex problems, such as problems routinely encountered in areas such as visual inspection of surfaces, rule-based solutions fall short.
Even relatively small changes, such as weaker printer ink, changing lighting conditions, or simply a change in supplier materials, can seriously impact the effectiveness of rule-based systems. Quality managers had to repeatedly call the service team to update parameters. It was clear that we needed more sophisticated technology to overcome these issues and be more forgiving in our applications. So deep he entered the age of learning.
The first era of deep learning
Deep learning is now the primary method of providing quality inspection. Deep learning solutions have proven to be able to solve more complex problems that are difficult to define with rule-based systems, such as detecting poorly defined surface scratches and dents. By receiving many examples of defects, the deep learning model can generalize the problem and develop this more general understanding to detect defects in new parts. However, there was a problem.
The process of providing these images to the model for training must be done manually. This is a very long process and not easy. During model training, we need to review many images. For example, if the model cannot detect cracks accurately, you may need to add more samples to the data until the model understands the cracks. Additionally, simply providing an image is not enough, you also need to provide markup around the image. For example, you may need to draw a box around the defect, or segmentation to trace the exact line of the crack.
Note that the model can only be trained on images of defects. End-users have to go through thousands of images, identify faulty images, and perform correct markup before serving them to the machine. Most deep learning applications require hundreds, if not thousands, of examples just to get started. If you’re making 100,000 parts a day and you know his 2% of the parts you make are defective, that’s a ton of images to review.
One approach to circumvent this problem is to feed the model with an artificial image of the defect. In other words, the customer manually creates defects and feeds these images into the model. Some companies are trying to develop software to generate images of potential defects, but both approaches face the same problem. Artificially created defects cannot accurately represent defects encountered in real-world situations and occurring in the natural production process.
Unsupervised Learning: A New Era?
We are now at the beginning of a new third era of quality inspection. Technical solutions aimed at overcoming the limitations of previous deep learning solutions are commonly referred to as unsupervised or semi-supervised systems. The main goal is to automate the model building process described above.
While traditional deep learning solutions require examples of defects to train, semi-supervised models can be fed images of samples of parts with no defects (“OK parts”). The first model, while not perfect, provides a basic understanding of what’s okay. Use this generalized understanding to flag suspicious defects or outliers and use customer feedback to continuously update and learn.
Automating the model building process in this way is much easier for the end user, but the greatest return on investment comes from reducing the time it takes to create a working model.
In this new era of unsupervised learning, the process of model building can be done on the production line. Previously, the data classification process required one person to work on her own for at least a week before taking the model back to the production line. After deploying a model, it often did not work as intended, often requiring further training or disruption of the production environment. In this new era of unsupervised or semi-supervised learning, the first training mode can be done on the production line and ready for deployment in as little as 24 hours.
A semi-supervised solution is arguably superior to a fully supervised system, as it results in a system that understands the product better. In the past, many of the people working on training models were data and AI experts, but they didn’t understand the product well enough. In this new era, semi-supervised systems leverage the knowledge of the producers and it is their feedback that helps optimize the model.
A final advantage for companies willing to embrace this new era is the ability to distinguish between types of defects. Previously, the only goal was to detect defects, and classification was less important. However, more advanced deep learning solutions can cluster different images together to build an understanding of different defect types. The data collected from here helps support both preventive and predictive maintenance.
In any area of technological progress, we are entering a new era that is not just a marginal change, but a paradigm shift. Breakthroughs in AI have made it possible to solve problems that were previously too complex for deep learning solutions, and new innovations have made the difficult task of building viable models faster and easier thanks to automation. The times are starting to come. Lean AI uses patented technology to develop semi-supervised solutions for quality inspection in various industrial applications.
For more information, please visit www.lean-ai-tech.com.
Homepage link
