Deep learning approaches for dislocation segmentation in TEM

Dislocations are defects of fundamental importance because their dynamics impacts alloys mechanical properties. At the TEM scale, dislocation dynamics analysis requires an important knowledge in image contrast and time consuming measurements. Here, we present several approaches using deep learning that aim at facilitating further analysis by performing dislocation segmentation in a large variety of materials and imaging conditions in TEM. We propose both fully supervised learning (FSL) and semi-supervised learning (SSL) approaches, using an encoder-decoder neural network architecture and a boundary-type loss. In FSL, training is performed on labeled images, while SSL uses in addition images in an unsupervised manner in order to enrich the feature description provided by an in-house large unlabeled dataset. We show that evaluation metrics are improved by SSL and almost reach the human expert performance. Finally, we explore the use of synthetic images for an unsupervised learning (domain adaptation). We constructed a dataset on physical grounds, using image simulation and dislocation dynamics simulation. Despite a reasonable good visual agreement, we show that feature knowledge is hardly transferred from synthetic to real images, leading to lower performances. However, synthetic images can be valuable to improve predictions in more difficult imaging conditions. A straightforward application example to automatic density measurement is presented as perspective.