COVID-19 Face Mask Recognition

Shahar Rotem - 206485898
Afik Bar - 311121289

Data

The dataset contains 24,345 facial image, splitted to train and test sets.

Figure 1. Train and Test Images samples, Respectively.

As we can see from this random samples, Images seems to cover a variety of ages (infants and seniors), and a multiple races.
Furthermore, we can notice that the masks "type" are not fixed to certain color or shape.
Both of the sets are balanced, as we can see from the following plots:

Figure 2. Train and Test Images Mask Ratios, Respectively.

Experiments

7 Layers CNN

Preprocessing

To obtain better generalization, we’ve used several transformations:

• Random Cropping

• Random Horizontal Flipping

• Normalizing each color channel.

Model Architecture

Our neural network consists of 7 Convolutional layers with batch normalization & ReLU as activation function, 1 average pool, and 1 fully connected layers.

Figure 3. CNN Architecture

Loss Function

We’ve used Cross Entropy loss function.

Optimizer

We’ve tried several different optimizers, with different configurations, and found out that Adam-Weighted performed slightly better, with default decay & learning rate.

Regularization

We found out that Dropout consistently preformed worse (Possibly due to Batch normalization).
Our optimizer utilizes Weight decay, which limits the learned change in weights for each batch, we’ve used the default value of 0.01.

Convergence

We were able to reach a F1-Score of 96.1 on the test-set.

Figure 4. 7 Layers CNN Convergence

Conclusions

We found out that architecture changes had slight improvements over the data, and most of the improvements resulted from data augmentation.

MobileNet

Preprocessing

To obtain better generalization, we’ve used several transformations:

• Random Cropping

• Random Horizontal Flipping

• Normalizing each color channel.

Model Architecture

We’ve used MobileNet V3.

Figure 5. Specification for MobileNetV3

Loss Function

We’ve used Cross Entropy loss function.

Optimizer

We’ve used Adam-Weighted Optimizer, along with Learning Rate Scheduler. Our initial learning rate was 0.001, and scheduler gamma is 0.1.

Regularization

We found out that Dropout didn’t affect results at all (Possibly due to Batch normalization).
Our optimizer utilizes Weight decay, which limits the learned change in weights for each batch, we’ve used the default value of 0.01.

Convergence

We were able to reach an accuracy of 97.95 on the test set.

Figure 6. MobileNet Convergence

Conclusions

We’ve tested multiple published models, however, MobileNet intrigued us the most, since it is intended to by used by low budget processing units (In smartphones), which we believe fit the task (: Side note - PyTorch has an implementation of MobileNetV2, but we’ve found out that V3 performed just slightly better.