Schematic drawing of the network architecture of the autoencoder (a) as well as the single-pixel-decoder (b) for 6.25% compression ratio. For lower compression ratios, a down/up-convolutional block as well as an encoder block was added. The encoder/decoder block consists of two convolutional layers with kernel size 3 and 128 filters each as well as a skip connection. As activation function we used LeakyRelu. The weights of the decoder part of the single-pixel-decoder are shared with the autoencoder network and not retrained. DownConv: Convolutional layer with stride 2, EncBlock/DecBlock: Encoder/Decoder block, MaxPool: Maximum pooling layer, UpConv: Convolutional layer followed by a transpose convolutional layer with stride 2, UpSamp: Upsampling layer.