Image Based Quality Control in Thermoplastic Composite. Production using Deep Transfer Learning
Bildbasierte Qualitätskontrolle in Thermoplastischen Verbundwerkstoffen. Produktion mittels Deep Transfer Lernen
A transfer learning approach for studying the classification of patterns in a UDtapecontext is presented. For this, we develop a convolution neural network(Surface-CNN), pre-train it on a surface dataset, and transfer the knowledge to other publiclyavailable datasets. The network was able to achieve 85% accuracy on the validationset. Next, we perform a comparative evaluation, by using state-of-the-art imageclassification networks. Here ResNet-50 and EfficientNet-B0 underperformed indifferent learning settings. To improve the performance of classification we use anensemble learning strategy which is to use model averaging by stacking. Finally, toimprove the scope of this study we perform a regularisation technique called EWCto achieve incremental learning for datasets arriving at d ifferent time intervals. Weprovide an effective transfer learning strategy to meet the needs of UD-Tape datasetanalysis and to have an effective classification of surface image datasets.
Sankt Augustin, Hochschule Bonn-Rhein-Sieg, Master Thesis, 2021