Evaluation of GAN for detection of anomalies through image analysis by cognitive cameras
| Reference number | |
| Coordinator | Gimic AB |
| Funding from Vinnova | SEK 300 000 |
| Project duration | November 2018 - July 2019 |
| Status | Completed |
| End-of-project report | 2018-02726sv.pdf(pdf, 515 kB) (In Swedish) |
Purpose and goal
The objective of the study was to evaluate alternative Machine Learning models that can be used for automated defect inspection with cameras in the manufacturing industry. We also wanted to implement UI support so that end users without knowledge about machine learning can create datasets and train models by themselves.
Expected results and effects
Several important results have been achieved in this project: Gimic has developed support in its platform to support different types of machine-learning models. Using the platform, a new model type based on "Siamese networks" has been implemented and evaluated. Our evaluation shows that this type of model is particularly well suited for use cases where the inspection object is fixed in a certain position. GUI support has also been developed where the end user can manage datasets and train models. This has proven to be very appreciated by the end users of the system.
Planned approach and implementation
The project started with a development phase, where the basic software components were designed and implemented. When this was done, a test station was set up at Gimic´s office where machine learning models were trained and compared. After an evaluation, software support was implemented to run models of the type "Siamese networks" on Gimic´s Edge platform. As a final step in the project, a prototype of the entire system was constructed. This prototype is planned to be installed and tested in the factory at Hordagruppen at the end of July 2019.