Faster Training of Deep Convolutional Neural Networks for Material Science through Transfer Learning

Convolutional neuronal networks (CNNs) have shown impressive results in the segmentation of biomedical data and more recently also of industrial data, especially of computed tomography (CT) data of fiber reinforced polymers (FRP). In these application areas, it is still a large challenge to generate ground truth data which is necessary for training respective CNNs. In addition, training a CNN from the ground up is a time-consuming process. This paper will show how transfer learning can be used to reduce the effort of training in the case of CT data segmentation. To achieve this, a pre-trained CNN is used, which allows to segment CT scans with distinct features of different grey values. To improve the prediction of the network based on the original training, only a small fraction of manually segmented reference data should be used. The proposed method is tested with CT data of a short glass fiber reinforced polymer. Due to the different absorption properties of the materials prevalent in such composite materials, the areas of the individual materials appear in different grey values. Between matrix and air/voids there is only a small difference in the absorption. With a threshold-based manual segmentation method, it is not possible to get proper results without splitting the volume into smaller regions of interest and update the thresholds for every region by hand. Our case studies will demonstrate that pre-trained CNNs have a high potential for partly replacing the time-consuming manual segmentation in the future. Through transfer learning, it was possible to reduce the amount of manually segmented regions significantly, without negatively impacting the segmentation accuracy.