Table 1 Feature map size and channel changes across different blocks of the proposed LMSAUnet architecture.

From: Lightweight Unet with depthwise separable convolution for skin lesion segmentation

Blocks

Output feature map size

Number of input channels

Number of output channels

ECDF BLOCK 1

H * W

3

64

DOWNSAMPLING 1

H/2 * W/2

64

64

ECDF BLOCK 2

H/2 * W/2

64

128

DOWNSAMPLING 2

H/4 * W/4

128

128

ECDF BLOCK 3

H/4 * W/4

128

256

DOWNSAMPLING 3

H/8 * W/8

256

256

ECDF BLOCK 4

H/8 * W/8

256

512

DOWNSAMPLING 4

H/16 * W/16

512

512

UPSAMPLING 1

H/8 * W/8

512

512

ECDF BLOCK 5

H/8 * W/8

1024

512

UPSAMPLING 2

H/4 * W/4

512

512

ECDF BLOCK 6

H/4 * W/4

768

256

UPSAMPLING 3

H/2 * W/2

256

256

ECDF BLOCK 7

H/2 * W/2

384

128

UPSAMPLING 4

H * W

128

128

ECDF BLOCK 8

H * W

192

64

PWCONV

H * W

64

2

  1. Each block illustrates the corresponding output resolution (H × W), input channels, and output channels at that stage.
Back to article page