A CNN architectures A.1 DnCNN

In this section we describe the denoising architectures used for our computational experiments. All architectures except BFCNN have additive (bias) terms after every convolutional layer. DnCNN [66] consists of 20 convolutional layers, each consisting of 3 3 filters and 64 channels, batch normalization [23], and a ReLU nonlinearity. It has a skip connection from the initial layer to the final layer, which has no nonlinear units. We use BFCNN [37] based on DnCNN architecture, i.e, we remove all sources of additive bias, including the mean parameter of the batch-normalization in every layer (note however that the scaling parameter is preserved). Our UNet model [50] has the following layers: 1. conv1 - Takes in input image and maps to 32 channels with 5 5 convolutional kernels. The input to this layer is the concatenation of the outputs of layer conv7 and conv2. The structure is the same as in [68]. This configuration of UNet assumes even width and height, so we remove one row or column from images in with odd height or width. We use a modified version of the blind-spot network architecture introduced in Ref. [29]. We rotate the input frames by multiples of 90 and process them through four separate branches (with shared weights) containing asymmetric convolutional filters that are vertically causal. The architecture of a branch is described in Table 1.