Free Access
Table 1.
Architecture of the encoder–decoder network.
| Layer | Type | Kernel size (a) | Stride | Input tensor shape(b) | Output tensor shape(b) |
|---|---|---|---|---|---|
 |
BN+ELU+CONV | 3 × 3 × C | 1 | W × W × C | W × W × 32 |
 |
BN+ELU+CONV | 3 × 3 × 32 | 2 | W × W × 32 | W/2 × W/2 × 32 |
 |
BN+ELU+CONV | 3 × 3 × 32 | 1 | W/2 × W/2 × 32 | W/2 × W/2 × 32 |
 |
BN+ELU+CONV | 3 × 3 × 32 | 2 | W/2 × W/2 × 32 | W/4 × W/4 × 32 |
 |
BN+ELU+CONV | 3 × 3 × 32 | 1 | W/4 × W/4 × 32 | W/4 × W/4 × 32 |
 |
BN+ELU+CONV | 3 × 3 × 32 | 2 | W/4 × W/4 × 32 | W/8 × W/8 × 32 |
 |
BN+ELU+CONV | 3 × 3 × 32 | 1 | W/8 × W/8 × 32 | W/8 × W/8 × 32 |
 |
CONV | W/8 × W/8 × 512 | 1 | W/8 × W/8 × 32 | 1 × 1 × 512 |
Notes. The naming convention for the convolutional blocks is Cb, l, with b referring to the label indicated above each block in Fig. 1 and l to the layer inside each block.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.