If an automated inspection system is unable to reliably distinguish between a microscopic defect on the surface of steel and a false alarm, does it truly solve the problem, or does it simply create a new and more costly one?
With this major challenge in mind, Josu Viteri has developed his Final Degree Project (TFG) in our Quality department. In this R&D project, a three-stage computer vision solution (information on the area of interest, refinement, and semantic segmentation of defects) is being designed and implemented for the quality control of steel surfaces in an industrial setting.
The key points of this development are:
- Architectural change: A legacy, unsupervised, low-performance process has been overhauled, and a supervised solution based on active learning has been designed in its place.
- Significant improvements in metrics: Accuracy has increased by 181% (from 0.32 to 0.90) and IoU by 178% (from 0.32 to 0.90), virtually eliminating costly false alarms while ensuring a recall of 98.4% (F2 score of 0.96).
- Implementation: The ROI detection model has been fine-tuned, and the system has been deployed for production use.
This major breakthrough in advanced analytics is being made possible at this stage of development thanks to Josu’s work and the close collaboration of the entire quality team: Ander Cazallas, Aitor Ciriaco, and Gartzen Olabarrieta are providing technical expertise and overseeing the results, under the direction and supervision of Goretti Frias.
We’re very proud to support young talent and to see university projects have this level of real impact on our innovation!
Thinking together
