A domestic research team has succeeded in developing a technology that can automatically recognize ultra-fine defects on surfaces with complex shapes.

The research team led by Dr. Kang Sung-hoon of the Materials Artificial Intelligence and Big Data Laboratory at the Korea Institute of Materials Science (KIMS, President Lee Jeong-hwan), a government-funded research institute under the Ministry of Science and ICT, has developed a technology that can accurately detect ultra-fine defects on the surface of products that are difficult for humans to distinguish using artificial intelligence (AI) technology.

Minor defects in products appear in various forms, and visually checking and inspecting them results in a lot of time and cost loss. In particular, the surface of parts with complex hierarchical structures has a three-dimensional shape, making it somewhat difficult to obtain accuracy using only a microscope or visual inspection.

To solve these problems, the research team succeeded in developing artificial intelligence (AI) technology that automatically recognizes the hierarchical structure of a product's surface through an image that simplifies a three-dimensional shape into a two-dimensional shape and detects micro-defects that deviate from the reference image.

The applied artificial neural network automatically learns the characteristics of defective factors and then performs defect detection and classification.

In particular, ultra-fine defects, which are half the thickness of a human hair (50㎛), are impossible to accurately read with the naked eye, so workers had to take dozens of photos of each specimen using a microscope and check them one by one. However, using this technology, it is possible to accurately detect defective factors in just one second or so.

In addition to defect detection, the Class Activation Map (CAM) neural network was applied to allow users to know which surface features the artificial intelligence (AI) used to select defects, thereby increasing the reliability of ultra-fine defect detection.

“Process automation and smart factory construction have recently become social issues,” said Kang Seong-hun, the research director and principal researcher. “If the developed technology is applied to various automated process lines, it is expected that even minute defects that are difficult for humans to read will be detectable in a short period of time.”

This research was conducted with support from the Ministry of Trade, Industry and Energy's 'Virtual Engineering Platform Construction Project for Lightweight Materials for Automotive' and the Korea Institute of Materials Science's major projects.