journal menu
![[Figure 2]](yr5131fig2mag.jpg)
|
|
Figure 2
Architecture of the DL model. It is based on a modified U-Net architecture with the use of dilated convolutions in the first two encoder blocks (left red circle) where the input was concatenated to its convolutions with different dilation rates (d parameter) before the MaxPooling layer. The gap-affected small portions P, in batches (B) of 32, were used as input (top left) and progressively sent through the encoder section down to a small volume of 2 × 2 × 2 pixels. Each building block of the decoder section (orange cubes) takes as input the concatenation of its previous block's output with the corresponding output of the encoder block of the the same size. The final output (top right) is a batch of inpainted P. |
![[Figure 2]](yr5131fig2.jpg)
 | JOURNAL OF APPLIED CRYSTALLOGRAPHY |
ISSN: 1600-5767
access
