What are the challenges for vision inspection in construction?
We mainly face three challenges at our projects:
1. The systems work in difficult environments, often dirty, rough and hard to access
This point does not need a lot of explanation. In most projects we have to deal with dust, water, or mud and cameras might be installed in very high positions, or very narrow ones or even underground or on flying objects.
2. There is only a low number of actual defect images per category to train models
For decades, the industry has focused on avoiding errors through process optimization. For computer vision, especially if Machine Learning (ML) is used, which learns strong models from large data sets, this is a challenge. Many of our customers lack a database of defective parts. Some may have 100, 10 or even fewer images of a defect they would like to detect.
3. There are quickly changing requirements and camera positions
With the constant improvement of products and processes, there is an increasing chance that vision inspection systems will quickly fail to meet decided and trained conditions. Product changes and changes in the production environment put inspection systems at risk and can even lead to the discontinuation of the original system.
How to design successful systems?
In the following we share some key learnings in designing and implementing a successful system to overcome the mentioned challenges.
a) Autonomous working and specially protected systems
Even though it is technically possible to use cloud computing for most of our applications we decided in all of our current projects to use edge devices. This allows us to avoid networking issues and guarantee a maximum reliability. Obviously, we also found ways to protect our hardware equipment according to the unique environment and fulfill the often very strict safety requirements (e.g. ATEX).
b) Be flexible in your model creation
Having few images of defects are actually one of the most complex challenge the full machine vision industry and research face. Probably we will discuss this issue one day in a special blog post. Anyway, there are some things we do to overcome such issues:
- Combine traditional methods with ML – many times traditional algorithms are more reliable and also cheaper to develop
- ML models which trains with “only good” data. We successfully refined unsupervised and semi-supervised algorithms for anomaly detection to target domains inside the construction industry. (this one is kind of a recipe secret of our company, which is why I won’t give too many details at the moment)
- Use cross-project datasets for model training
- If the upper points don’t work, go the extra mile to collect high quality data. Once we literally traveled through Europe to create a dataset.
c) Be flexible in error definitions and expect changes in environment
In production, conditions are adjusted. An accepted tolerance for a crack width of 0.4 mm may be 0.2 mm the next month. Further, the definition of a defect can be very subjective and it is common for disagreements to arise between two inspectors. Therefore, our annotation and software development processes in cooperate and expect changes in environment and specifications.
Vision inspection systems can dramatically simplify and improve quality processes in the construction industry. Herein, ML is not the solution to all problems, often traditional methods are still the best option to solve a problem. However, there are various uses of ML to make processes more effective, autonomous and cheaper. As in any project, it pays out to define the specification clearly and to research and weigh problem solutions well. By asking the right questions and using the right tools in production, it is possible to create value with ML in even harsh environments such as the construction industry and make your processes more efficient and better.
preML GmbH is a computer vision start up that develops customer-specific solutions in the field of image processing and object detection for industry and the construction sector. One focus is on the detection and monitoring of building elements, especially those made of concrete, with the help of camera technology and intelligent algorithms. The start-up was founded in 2020 at KIT from the Department of Computer Science, and the company’s current headquarters are in Lahr in the Black Forest.