Application of Transfer Learning for Detecting Fiber Orientations in Images of Fiber Reinforced Plastics

Abstract

Fiber reinforced plastics are an essential material for lightweight products. However, their superior mechanical properties compared to traditional materials are only guaranteed, if necessary quality requirements are met (e.g. fiber orientation). One promising approach for detecting quality deviations in image-based monitoring data is to use deep learning models. Nevertheless, these models need vast amounts of labeled training data, which is either not available or expensive to attain. To train deep learning models more data efficiently, a common and intuitive transfer learning approach is applied to detect fiber orientations for two different fiber reinforced plastics. By comparing the similarity between different domains of public datasets with the domain of the materials of this work, the range of optimal hyperparameters were estimated for the transfer learning task at hand. Through a grid search within the estimated hyperparameter range, the best-performing models were identified, showing that models transferred from similar domains do not only result in better performance but are also more robust against data scarcity. Finally, the results show that transfer learning holds the potential to accelerate the usage of deep learning for quality assurance tasks in textile-based manufacturing.