Fig. 2

Network models. Image data containing the temporal context of the previous and subsequent frame are processed in different network models. a CS CNN uses compressed sensing as a prior and applies several convolutional layers. b CS Inception integrates the CS component deeper into the neural network. c CS U-Net uses compressed sensing as a prior and computes the feature space with a U-Net architecture d Rec U-Net aims to unroll the CS algorithm with iterative encoding and decoding from image to feature space and vice versa. For all network models the feature space is processed with sigmoid and tanh activations and fed into a Gaussian mixture model to compute the loss