Open Access
Table B.1.
CNN architecture.
| Layer type | # Param. | Output shape | Properties |
|---|---|---|---|
| Input | 0 | (1,N,N) | |
| Convolutional | 1600 | (32,N,N) | 1 pixels stride, |
| 32 filters (7,7) | “same” padding, | ||
| ReLU act. | |||
| Max Pooling | 0 | (32,N/2,N/2) | pool size 2 |
| Dropout | 0 | (32,N/2,N/2) | 25% |
| Convolutional | 100416 | (64,N/2,N/2) | 1 pixels stride, |
| 64 filters (7,7) | “same” padding, | ||
| ReLU act. | |||
| Max Pooling | 0 | (64,N/4,N/4) | pool size 2 |
| Dropout | 0 | (64,N/4,N/4) | 30% |
| Batch Norm. | 256 | (64,N/4,N/4) | |
| Convolutional | 401536 | (128,N/4,N/4) | 1 pixels stride, |
| 128 filters (7,7) | “same” padding, | ||
| ReLU act. | |||
| Max Pooling | 0 | (128,N/8,N/8) | pool size 2 |
| Dropout | 0 | (128,N/8,N/8) | 30% |
| Convolutional | 802944 | (128,N/8,N/8) | 1 pixels stride, |
| 128 filters (7,7) | “same” padding, | ||
| ReLU act. | |||
| Max Pooling | 0 | (128,N/16,N/16) | pool size 2 |
| Dropout | 0 | (128,N/16,N/16) | 30% |
| Flatten | 0 | (N2/2) | |
| Dense | (N2/2+1) | (512) | 512 units, |
| ×512 | ReLU act. | ||
| Dropout | 0 | (512) | 35% |
| Dense | 65664 | (128) | 128 units, |
| ReLU act. | |||
| Dropout | 0 | (128) | 35% |
| Dense | 129 | (1) | 1 unit, |
| sigmoid act. | |||
Notes. Columns are the name of the Keras layer, the number of trainable parameters, output, and hyper-parameters for each layer. N is the size of the input images (256 / 64 for the original resolution / “degraded” images). ReLU activation function stands for Rectified Linear Unit.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.