Using Artificial Intelligence to Detect Defects on Surfaces Better and Faster

BMBF Project »SynosIs« – Synthetic, Optically Realistic Image Data of Surface Structures for AI-Based Inspection Systems

In the project »SynosIs«, we are working with our research partners to develop an inspection system based on artificial intelligence (AI) that detects defects on surfaces quickly and automatically. The project is funded by the German Federal Ministry of Education and Research (BMBF).

Image processing systems for quality control in production are solving increasingly complex tasks and, due to short development cycles, must react ever more quickly to new products and defect patterns. Typical inspection tasks include the detection of surface defects and deviations from the target geometry. Artificial intelligence (AI) is used successfully in image recognition, processing and understanding to some extent. However, training a robust, automatic AI-based inspection system requires a large amount of manually annotated image data, especially one that is representative of all defect types.


Challenges of an AI-Based Inspection System:

  1. Many defects occur very infrequently. In particular, defects that are critical to component safety are observed less frequently because they are avoided as much as possible on the production side.
  2. Manual annotation of exemplary defects is therefore not only time-consuming, but also leads to inconsistent results in some cases.
  3. In particular, the decisions of the trained system may be biased by the lack of defect images in the training.

It is therefore desirable to automatically generate realistic synthetic image data together with the corresponding annotations. Currently, the simulation of such image data requires multiple manual interactions. Moreover, synthetic generation of surface defects is an open research topic.

Frequently Asked Questions

Why is it so difficult to train AI to reliably inspect a machined surface?

Because the texture variations and features often lead the AI to suspect a defect where there is none, or to mask the defects in certain situations.

Why don't we train the network with more defect examples to make it more robust?

Because obtaining and annotating such a large amount of data is usually expensive and impractical.

Why don't we just use correct surfaces as training?

Conspicuity detection algorithms are not yet able to handle the variation caused by complex textures and focus on the often minor changes caused by surface defects.

Can we use artificially generated images to support AI?

Yes, this is also the latest state of development. There are a lot of open questions, and we are addressing many of them with our SynosIs software. Our core motivation here is that no two produced objects are exactly alike, and the image simulation pipeline needs to be able to reproduce that exact idea. It should also retain control over the key production parameters.

Privacy warning

With the click on the play button an external video from is loaded and started. Your data is possible transferred and stored to third party. Do not start the video if you disagree. Find more about the youtube privacy statement under the following link:

The SynosIs »Pipeline«

Approaching Reality Automatically With AI and Building up an Image Data Portfolio

In this interdisciplinary project, methods from physics, mathematics and computer science are combined to obtain synthetic images of typical defects on metallic surfaces – at a level of realism not currently achieved. This image data is openly accessible and can be used for training and validating AI models. This simplifies and accelerates the development of AI models for optical surface inspection.

For the first time, the image data portfolio will provide guaranteed correct and objectively annotated defect images. Furthermore, this synthetic data enables an objective quantitative comparison of different solutions as well as the prediction of the probability of detection (POD) of surface defects depending on defect, component and inspection system.

Our Expertise: Define Defects, Classify, Train AI

In the initial stage, our ITWM experts, in close cooperation with Fraunhofer IOF, design two free-form specimens with complex geometry for validation. The defects to be introduced are defined in such a way that, on the one hand, typical defect classes such as impact points, scratches, and (oil) contamination are represented, and, on the other hand, the manufacturing effort remains reasonable. In the next step, the researchers train the AI-based inspection system using synthetic image data. Another task is to validate the inspection solution and simulate a representative portfolio of synthetic image data in collaboration with the Rhineland-Palatinate Technical University Kaiserslautern-Landau (RPTU).

This project has a clear application scenario, but is still designed as a basic research project because we are developing new methods in all three disciplines. In addition to data synthesis, the project addresses the topic of »parameters for measuring suitability, quality or bias-freeness of the data«, since suitability, quality and bias freedom of the data are continuously checked and guaranteed for both the defect-free surfaces and the defects to be introduced.

Current Status in the »SynosIs« Project

In the first year of »SynosIs«, we as a team realized that there are too few controlled data sets and modeling processes for physical surfaces. We therefore developed a process chain in the project to meet the need. In doing so, the team designed a dataset of ten aluminum test specimens that focuses on difficult surfaces that contain textures created by spiral milling, parallel milling, or sandblasting.

Based on topology measurements of the manufactured surfaces, we created models capable of copying the appearance of the surfaces using stochastic geometry modeling techniques. In this way, we digitally create an arbitrary number of surfaces that nevertheless have different appearances while maintaining the same properties. Moreover, we thus control the models by production-related parameters that influence the final surface.  

The surface generation models developed can be used again and again and combined to apply to any 3D model.

In the second year of the project, the team will focus on:

  • modeling of defects
  • creating public data sets
  • investigating the benefits and challenges of simulation-based image processing