Table 2.
Network structure of the residual network and spatial attention module.
| Unit | Layer | Filter/Stride | Output Size | |
|---|---|---|---|---|
| Input | 0 | 112 × 112 × 3 | ||
| Residual Network |
1 | Conv-BN-ReLU | 7 × 7, 64/2 | 56 × 56 × 64 |
| Max Pooling | 3 × 3/2 | 28 × 28 × 64 | ||
| 2 | Conv-BN-ReLU | 3 × 3, 64/1 | 28 × 28 × 64 | |
| Conv-BN | 3 × 3, 64/1 | 28 × 28 × 64 | ||
| 3 | Conv-BN | 1 × 1, 128/2 | 14 × 14 × 128 | |
| Conv-BN-ReLU | 3 × 3, 128/1 | 14 × 14 × 128 | ||
| Conv-BN | 3 × 3, 128/1 | 14 × 14 × 128 | ||
| 4 | Conv-BN | 1 × 1, 256/2 | 7 × 7 × 256 | |
| Conv-BN-ReLU | 3 × 3, 256/1 | 7 × 7 × 256 | ||
| Conv-BN | 3 × 3, 256/1 | 7 × 7 × 256 | ||
| 5 | Conv-BN | 1 × 1, 512/2 | 4 × 4 × 512 | |
| Conv-BN-ReLU | 3 × 3, 512/1 | 4 × 4 × 512 | ||
| Conv-BN | 3 × 3, 512/1 | 4 × 4 × 512 | ||
| Spatial Attention Module |
6 | AvgPool | 4 × 4 × 1 | |
| MaxPool | 4 × 4 × 1 | |||
| AvgPool + MaxPool | 4 × 4 × 2 | |||
| Conv-Sigmoid | 7 × 7, 1/1 | 4 × 4 × 1 | ||
| 7 | Product (5 ∘ 6) | 4 × 4 × 512 | ||
| Output | 8 | GlobalAvgPool | 512 |
BN: batch normalization. In Unit 7, the outputs of Units 5 and 6 are multiplied for each position in the feature maps.