Figure 3

Architecture of the convolutional neural networks designed for the segmentation of SEM and TEM images. For the SEM model, 3 convolutional layers are used at each block, while only 2 convolutional layers are used for the TEM model. Convolutional layers in dashed lines are removed for the TEM model. All activation functions used are rectified linear units (ReLU). Strided convolutions are used to downsample the features during the contraction path (left), while up-convolutions are used to recover the localization during the expansion path (right). Features of the contraction path are merged with features of the expansion path to combine localization and context (illustrated by the concatenation step). The pixel-wise classification is done by a 3-class softmax.