Table 1.
Architecture of feature classification module.
| Layer | Input | Filter | Kernel | Output |
|---|---|---|---|---|
| Conv2D | (H, W, 1) | F1 | (1, K1) | (H, W, F1) |
| BatchNorm | (H, W, F1) | (H, W, F1) | ||
| DepthwiseConv2D | (H, W, F1) | F1*D | (H, 1) | (1, W, F1*D) |
| BatchNorm | (1, W, F1*D) | (1, W, F1*D) | ||
| ELU activation | (1, W, F1*D) | (1, W, F1*D) | ||
| AveragePooling2D | (1, W, F1*D) | (1, P1) | (1, W/P1, F1*D) | |
| SeparableConv2D | (1, W/P1, F1*D) | F 2 | (1, K2) | (1, W/P1, F2) |
| BatchNorm | (1, W/P1, F2) | (1, W/P1, F2) | ||
| ELU activation | (1, W/P1, F2) | (1, W/P1, F2) | ||
| AveragePooling2D | (1, W/P1, F2) | (1, P2) | (1, W/(P1*P2), F2) | |
| Flatten | (1, W/(P1*P2), F2) | (W*F2)/(P1*P2) | ||
| Dense | (W*F2)/(P1*P2) | N |