Table 1.
MSI-MFNet architecture. Note that each “Conv” layer shown in the table corresponds to the sequence BN-ReLU-convolution. The blocks are as follows: MSI: multi scale input; I: input; DB-x: depth block (1–4), MF: multi feature and O: output.
| Block | Layers | Output Size H × W | Parameters | Repetition | # of Param. | % of Param. |
|---|---|---|---|---|---|---|
| MSI | Input-1 | a × a | Upscale 1x | - | - | - |
| Input-2 | b × b | Upscale 0.5x | - | - | - | |
| Input-3 | c × c | Upscale 0.33x | - | - | - | |
| Input-4 | d × d | Upscale 0.25x | - | - | - | |
| I | Concat MSI (1–4) | 224 × 224 | - | - | - | - |
| ZP | 230 × 230 | 1 × 1 | - | - | - | |
| Convolution | 112 × 112 | 7 × 7, Stride 2 | - | 9408 | 0.1 | |
| BN | 112 × 112 | - | - | 256 | - | |
| Activation | 112 × 112 | ReLU | - | - | - | |
| ZP | 114 × 114 | 1 × 1 | - | - | - | |
| MP | 56 × 56 | 3 × 3, Stride 2 | - | - | - | |
| DB-1 | Conv | 56 × 56 | 1 × 1 | × 6 | 338,304 | 1.8 |
| Conv | 3 × 3 | |||||
| Conv | 56 × 56 | 1 × 1 | - | 33,792 | 0.2 | |
| AP | 28 × 28 | 2 × 2, Stride 2 | - | - | - | |
| GAP-1 | 128 | - | - | - | - | |
| DB-2 | Conv | 28 × 28 | 1 × 1 | × 12 | 930,048 | 4.9 |
| Conv | 3 × 3 | |||||
| Conv | 28 × 28 | 1 × 1 | - | 133,120 | 0.7 | |
| AP | 14 × 14 | 2 × 2, Stride 2 | - | - | - | |
| GAP-2 | 256 | - | - | - | - | |
| DB-3 | Conv | 14 × 14 | 1 × 1 | × 48 | 8,180,736 | 43.3 |
| Conv | 3 × 3 | |||||
| Conv | 14 × 14 | 1 × 1 | - | 1,612,800 | 8.8 | |
| AP | 7 × 7 | 2 × 2, Stride 2 | - | - | - | |
| GAP-3 | 896 | - | - | - | - | |
| DB-4 | Conv | 7 × 7 | 1 × 1 | × 32 | 7,083,520 | 37.3 |
| Conv | 3 × 3 | |||||
| GAP-4 | 1920 | - | - | - | - | |
| MF | Concat GAP (1–4) | 3200 | - | - | - | - |
| O | Dropout | 3200 | - | - | - | - |
| BN | 3200 | - | - | 12,800 | 0.1 | |
| Softmax | 2/4/8 | - | - | 25,608 | 0.1 |