| Autor: |
Bohlig, D., Leutert, F., Kempf, F., Möschwitzer, D., Schilling, K. |
| Jazyk: |
angličtina |
| Rok vydání: |
2022 |
| Předmět: |
|
| DOI: |
10.25967/570347 |
| Popis: |
In the move towards image processing for Automated Integration and Testing (AIT) we explored defect detection on reflowed satellite PCBs using deep learning. To this end we utilized a convolutional neural network for semantic segmentation and subsequent instance segmentation for the detection of surface mounted devices, pollution and defects. A dataset of 16k labeled instances including devices, solder connections, solderballs, bridges and tombstoned components was created from our satellite fleet PCBs to train the network. The images were recorded using microscopes and industrial cameras and labeled using an active learning approach with human experts annotating the initial data. Then, the partially trained network labeled additional data with experts supervising the process and correcting predictions where necessary.We explored k-fold cross validation as well as dropout based uncertainty estimation for the prediction of samples that meaningfully extend our training data. Further we evaluated the benefits of the implemented procedures and the annotation speedup from the network assisted annotation. The resulting inspection system was successfully integrated into a human-robot collaborative workspace to increase its production efficiency. |
| Databáze: |
OpenAIRE |
| Externí odkaz: |
|
