Fig. 4.

Detailed network structure of res-unet, conv-m, and conv-s shown schematically in Fig. 3. The res-unet is a residual U-Net with four downsamplings and four upsamplings which consists of the residual blocks (each has three 3 × 3 convolutions, a parametric ReLU, and a batch normalization). Two output branches (conv-m, conv-s) consist of a residual block and a 1 × 1 final output layer.