CNN Models
U-net
The U-Net.
layers |
input |
output |
|---|---|---|
conv1 |
(B, H, W, 1) |
(B, H, W, cs) |
conv2 |
(B, H, W, cs) |
(B, H, W, cs) |
maxpl1 |
(B, H, W, cs) |
(B, H/2, W/2, cs) |
conv3 |
(B, H/2, W/2, cs) |
(B, H/2, W/2, 2cs) |
conv4 |
(B, H/2, W/2, 2cs) |
(B, H/2, W/2, 2cs) |
… |
… |
… |
Note
B: batch size
cls: number of class
cs: number of convolution kernel per layers
LRCS-Net
The LRCS-Net is a trimed model derived from Seg-Net.
Encoder:
layers |
input |
output |
|---|---|---|
conv1 |
(B, H, W, 1) |
(B, H, W, cs) |
maxpl1 |
(B, H, W, cs) |
(B, H/2, W/2, cs) |
conv2 |
(B, H/2, W/2, cs) |
(B, H/2, W/2, 2cs) |
maxpl2 |
(B, H/2, W/2, 2cs) |
(B, H/4, W/4, 2cs) |
conv3 |
(B, H/4, W/4, 2cs) |
(B, H/4, W/4, 4cs) |
maxpl3 |
(B, H/4, W/4, 4cs) |
(B, H/8, W/8 4cs) |
conv4sigm |
(B, H/8, W/8, 4cs) |
(B, H/8, W/8, 4cs) |
Decoder:
layers |
input |
output |
|---|---|---|
conv5a |
(B, H/8, W/8, 4cs) |
(B, H/8, W/8, 4cs) |
conv5b |
(B, H/8, W/8, 4cs) |
(B, H/8, W/8, 4cs) |
up1 |
(B, H/8, W/8, 4cs) |
(B, H/4, W/4, 4cs) |
conv6a |
(B, H/4, W/4, 4cs) |
(B, H/4, W/4, 2cs) |
conv6b |
(B, H/4, W/4, 2cs) |
(B, H/4, W/4, 2cs) |
up2 |
(B, H/4, W/4, 2cs) |
(B, H/2, W/2, 2cs) |
conv7a |
(B, H/2, W/2, 2cs) |
(B, H/2, W/2, 2cs) |
conv7b |
(B, H/2, W/2, 2cs) |
(B, H/2, W/2, 2cs) |
up3 |
(B, H/2, W/2, 2cs) |
(B, H, W, 2cs) |
conv8a |
(B, H, W, 2cs) |
(B, H, W, cs) |
conv8b |
(B, H, W, cs) |
(B, H, W, cs) |
conv8c |
(B, H, W, cs) |
(B, H, W, cls) |
Note
B: batch size
cls: number of class
cs: number of convolution kernel per layers