TABLE 3.
CNN architecture
| Layer type | Details |
|---|---|
| Input | Size 256 × 256 × 1 |
| Conv2D | Size 32 × 7 × 7, Stride 2 |
| MaxPooling2D | Size 2 |
| Conv2D | Size 64 × 5 × 5, Stride 1 |
| MaxPooling2D | Size 2 |
| Conv2D | Size 128 × 3 × 3, Stride 1 |
| MaxPooling2D | Size 2 |
| Conv2D | Size 256 × 3 × 3, Stride 1 |
| MaxPooling2D | Size 2 |
| Conv2D | Size 512 × 3 × 3, Stride 1 |
| Flatten | (None) |
| Dense | 256 Units |
| Dropout | 0.5 |
| Dense | 4 Units |
L2 kernel and activity regularization (1e−06, with default biases turned off) were applied to each Conv2D layer, with batch normalization (momentum = 0.01) applied between the Conv2D and MaxPooling2D layers. ReLU activation was used for all Conv2D layers and the first Dense layer, with Softmax activation applied to the output layer.