Fraunhofer Institute for Industrial Mathematics ITWM
Fraunhofer Institute for Industrial Mathematics ITWM
Anwendungsbeispiel Holzplatten: Maschinelles Lernen in der Oberflächeninspektion
© iStockphoto / Fraunhofer ITWM
Anwendungsbeispiel Holzplatten: Maschinelles Lernen in der Oberflächeninspektion

© iStockphoto / Fraunhofer ITWM

Application Example Wood Panels: Machine Learning in Surface Inspection

Optimize System for Inline Surface Inspection and Quality Control With Hybrid Approaches

Our researchers are developing new methods to automate the detection of defects on surfaces in production. We use hybrid methods that combine classic image processing with Artificial Intelligence (AI). This allows us to create a reliable and adaptable inspection system that works under real production conditions and significantly improves quality assurance. The systems inspect surfaces very accurately and find even small defects that are difficult to detect with the human eye.

We are working on making Machine Learning – especially Deep Learning – accessible for surface inspection. We develop and implement complex inline inspection systems based on Deep Neural Networks (DNN). One specific application example is the inspection system for wooden shelves, which uses various image acquisition techniques to achieve a high defect detection rate and replace manual inspection.

Classic Image Processing Methods and Modern AI-Based Processes: Hybrid Approaches as a Combination of the Best of Both Worlds

In many industrial production facilities, manual inspection by humans is still the method of choice when it comes to surface inspection. The reason for this is the flexibility of the workers. However, automated systems are more repeatable, the results are easier to track and they are more cost-effective in the long term.

While recognition systems modeled by experts were the standard for a long time, AI-based systems (also known as deep learning or data-based methods) have been increasingly widespread for around 15 years. These data-based methods are particularly unsuitable for inspecting natural materials such as wood.

However, the use of deep learning in complex online surface inspection systems is complicated. In particular, the »skewed« data situation is often a problem: there are usually few defect images and many »good« product parts. Furthermore, examples of all defect classes have to be collected in advance and their classification is difficult to change.

For this reason, we generally use a hybrid approach. Here we combine the advantages of classic image processing with those of deep learning methods. With the help of classic solutions, we first arrive at a quick initial solution, which we then use to generate a database for the AI-based methods.

Defect detection comparison: initial candidate search with classic image processing (left) and in the final system using AI (right).
© Fraunhofer ITWM
Defect detection comparison: initial candidate search with classic image processing (left) and in the final system using AI (right).
Custom Annotation Tool
© Fraunhofer ITWM
Custom annotation tool. As detailed as possible representation of the different images of the defect candidates (on the left among each other, normalized and in the original) as well as their position in the wood panel on the right. Center: Selection options for the annotation.

Practical Example: Surface Inspection System for Shelf Panels

An application example from our research practice is the inline inspection system for shelf panels. In future, it will replace manual inspection in the production of wooden panels. We use various image acquisition techniques to ensure a high defect detection rate and to meet the increased production requirements.

The product variants consist of single-colored panels and wood decor. The panels have different sizes. The defect catalog consists of typical defects such as scratches, holes or some specific defects such as discoloration, excess glue or lines. In particular, the wood-like decors with irregular patterns on their surfaces are a decisive reason to work with a deep learning algorithm here.

Experience: Effective Use of Deep Learning in Industrial Surface Inspection

The precise creation of the database is the key success factor for AI-based processes. Although the defect catalog serves as the basis for this, close communication between production, manual-visual inspection and the AI experts is essential in order to create this database. All parties involved have to translate their respective expertise.

This type of close cooperation was successful in detecting defects on wooden panels and a good database was found. The system achieves high precision and recall values for various defect categories. Our hybrid approach has shortened the development time and simplified communication with the various interested parties and the creation of the defect catalog. In this way, we can increase efficiency and establish reliable and consistent inspection systems. 

Customer defined defect type »sinkhole« with different intensities
Illustration of the variance of the error types: From left to right, the same error type from strong to weak with varying intensity.

Related Links: