Fig. 8: Architecture of the U-Net used in this work.

Here, the notation of Tensorflow⁵⁵ is adopted for naming the layers. The layers consist of separable convolutional layers (SeparableConv2D) with either a rectified linear unit (ReLu) or sigmoid activation functions to extract the features and apply the non-linearities, batch normalization (BatchNormalization) to transform the layers’ outputs to a mean value of zero and a standard deviation of 1, max pooling (MaxPooling2D) for coarse-graining, and up sampling (UpSampling2D) for going from the coarse grained image to a high resolution one. The skip connections send the images from each contracting step to its expanding counterpart.