阶段

输出 尺寸

卷积NeXt-SimAM分支

多尺度特征融合分支

Swin Transformer分支

-

32 × 32, 96

4 × 4, 96

-

4 × 4, 96

阶段一

32 × 32, 96

d7 × 7,961 × 1,3841 × 1,96

× 2d7 × 7,961 × 1,3841 × 1,96 × 2

→ 1 × 1, 96← → 1 × 1, 96←

Window size = 7 × 7 head = 3,1 × 1,96 × 2window size = 7 × 7 head = 3,1 × 1,96 × 2

阶段二

16 × 16, 192

2 × 2,192

1 × 1, 384

Patch merging

d7 × 7,1921 × 1,7681 × 1,192 × 2d7 × 7,1921 × 1,7681 × 1,192 × 2

Avgpool k2, s4

Window size = 7 × 7 head = 6,1 × 1,192 × 2 window size = 7 × 7 head = 6,1 × 1,192 × 2

阶段三

8 × 8, 384

2 × 2,384

→ 1 × 1,384← → 1 × 1,384←

Patch merging

d7 × 7,3841 × 1,15361 × 1,384 × 2d7 × 7,3841 × 1,15361 × 1,384 × 2

Window size = 7 × 7 head = 12,1 × 1,384 × 2 window size = 7 × 7 head = 12,1 × 1,384 × 2

分类器

1 × 1, 1

-

global average pooling

-

1 × 1, numclass