SupplementaryMaterialfor Uncertainty-DrivenLossforSingleImage Super-Resolution

Overall,twelvevarious monotonically increasing functions have achieved better results than the baseline method. The training cost mainly depends on the original networks and iterative times. Then the computational cost of the backpropagation can be approximately equal ton. Step2 training process is trained with4 105 minibatch updates and the variance branch doesn't require backpropagation. Figure 1 and Figure 1 shows the uncertaintyθ captured by proposedLUDL on three different SISR networks.