Table 1.
The whole framework of Dense MobileNetV3.
Fig 1.
Depthwise separable convolution.
(a) Depthwise convolution. (b) Pointwise convolution.
Fig 2.
Squeeze and Excitation (SE) module.
After excitation, different colors mean channels get different weights.
Fig 3.
The dark green area on the right is the detailed process of Dense Block.
Fig 4.
(a)Initial network. (b) Compact network. The compact network after pruning is fine-tuned to reach similar (or even better) accuracy than when trained normally.
Table 2.
Specific division of the database.
Fig 5.
Some cases of CXR from the COVID-19 Radiography Database.
(A) COVID-19 sample, (B) Pneumonia sample, and (C) Normal sample.
Table 3.
Hyper-parameter used during training.
Fig 6.
The fluctuation of accuracy and loss over epochs during training and validation.
Fig 7.
Confusion matrices distribution.
Table 4.
Ablation study metrics.
Table 5.
Comparison of the effect after using various pruning methods.
Table 6.
Comparison of parameters before and after pruning.
Table 7.
Dataset of five-fold cross-validation.
Table 8.
Results of five-fold cross-validation.
Table 9.
Evaluate accuracy and parameters by comparing the relevant model.
Fig 8.
F1-score comparison.
Fig 9.
Visualization of chest X-ray images using Grad-CAM on Dense MobileNetV3 model.