Mask-Detection

Mask Detection Program using Tensorflow, Keras and OpenCV. The Base Model being used here is ResNet10.

About

It is a Mask Detection Program to make people aware of wearing masks and protect themselves during the Corona pandemic.

Output

Without Mask @400p :


With Mask @400p :

Sample Output 1 :

Sample Output 2 :

Sample Output 3 :

All the section wise output files according to sections are in /output.

Dependencies

Install the dependencies from your terminal with -

pip install -r requirements.txt

Run

Run the Mask-Detector from your terminal with -

python detectMaskLive.py

Building the Model

*** Model Training ***

1. Firstly we need to create a Neural Network to classify the frames as positive/negative (Wearing a Mask or not).
2. We train a model using ResNet10 as the BaseModel and perform Data Augmentation to get more images from the given /dataset.
3. After creation of the NN we train the model with some given pre-defined hyperparamters. After which we get something like this -

*** Loading Models ***
*** Compiling our Model ***
*** Training our Model ***
Epoch 1/20
 1/34 [..............................] - ETA: 0s - loss: 1.1101 - accuracy: 0.46882020-07-11 21:35:10.388733: W tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 154140672 exceeds 10% of free system memory.
 2/34 [>.............................] - ETA: 15s - loss: 1.0499 - accuracy: 0.51562020-07-11 21:35:11.391032: W tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 154140672 exceeds 10% of free system memory.
34/34 [==============================] - 49s 1s/step - loss: 0.4660 - accuracy: 0.7940 - val_loss: 0.1289 - val_accuracy: 0.9819
Epoch 2/20
34/34 [==============================] - 52s 2s/step - loss: 0.1274 - accuracy: 0.9579 - val_loss: 0.0597 - val_accuracy: 0.9855
Epoch 3/20
34/34 [==============================] - 51s 1s/step - loss: 0.0742 - accuracy: 0.9794 - val_loss: 0.0443 - val_accuracy: 0.9855
Epoch 4/20
34/34 [==============================] - 51s 1s/step - loss: 0.0606 - accuracy: 0.9785 - val_loss: 0.0328 - val_accuracy: 0.9928
Epoch 5/20
34/34 [==============================] - 48s 1s/step - loss: 0.0449 - accuracy: 0.9888 - val_loss: 0.0292 - val_accuracy: 0.9891
Epoch 6/20
34/34 [==============================] - 51s 1s/step - loss: 0.0351 - accuracy: 0.9860 - val_loss: 0.0267 - val_accuracy: 0.9928
Epoch 7/20
34/34 [==============================] - 52s 2s/step - loss: 0.0333 - accuracy: 0.9888 - val_loss: 0.0250 - val_accuracy: 0.9928
Epoch 8/20
34/34 [==============================] - 51s 2s/step - loss: 0.0293 - accuracy: 0.9925 - val_loss: 0.0206 - val_accuracy: 0.9928
Epoch 9/20
34/34 [==============================] - 49s 1s/step - loss: 0.0295 - accuracy: 0.9899 - val_loss: 0.0240 - val_accuracy: 0.9928
Epoch 10/20
34/34 [==============================] - 49s 1s/step - loss: 0.0224 - accuracy: 0.9925 - val_loss: 0.0217 - val_accuracy: 0.9928
Epoch 11/20
34/34 [==============================] - 50s 1s/step - loss: 0.0200 - accuracy: 0.9934 - val_loss: 0.0199 - val_accuracy: 0.9928
Epoch 12/20
34/34 [==============================] - 50s 1s/step - loss: 0.0245 - accuracy: 0.9934 - val_loss: 0.0205 - val_accuracy: 0.9928
Epoch 13/20
34/34 [==============================] - 49s 1s/step - loss: 0.0200 - accuracy: 0.9925 - val_loss: 0.0197 - val_accuracy: 0.9928
Epoch 14/20
34/34 [==============================] - 47s 1s/step - loss: 0.0165 - accuracy: 0.9944 - val_loss: 0.0173 - val_accuracy: 0.9928
Epoch 15/20
34/34 [==============================] - 41s 1s/step - loss: 0.0156 - accuracy: 0.9963 - val_loss: 0.0176 - val_accuracy: 0.9928
Epoch 16/20
34/34 [==============================] - 40s 1s/step - loss: 0.0151 - accuracy: 0.9972 - val_loss: 0.0150 - val_accuracy: 0.9964
Epoch 17/20
34/34 [==============================] - 41s 1s/step - loss: 0.0118 - accuracy: 0.9981 - val_loss: 0.0147 - val_accuracy: 0.9928
Epoch 18/20
34/34 [==============================] - 41s 1s/step - loss: 0.0121 - accuracy: 0.9972 - val_loss: 0.0155 - val_accuracy: 0.9928
Epoch 19/20
34/34 [==============================] - 40s 1s/step - loss: 0.0113 - accuracy: 0.9972 - val_loss: 0.0135 - val_accuracy: 0.9964
Epoch 20/20
34/34 [==============================] - 42s 1s/step - loss: 0.0063 - accuracy: 1.0000 - val_loss: 0.0163 - val_accuracy: 0.9928
*** Evaluting our NN ***
              precision    recall  f1-score   support

   with_mask       1.00      0.99      0.99       138
without_mask       0.99      1.00      0.99       138

    accuracy                           0.99       276
   macro avg       0.99      0.99      0.99       276
weighted avg       0.99      0.99      0.99       276

*** Saving Mask-Detector Model ***

4. Now our Model has been saved as maskDetector.model which is a h5 type file.
   
5. All the above steps can be performed by running python trainMaskDetector.py -d ./dataset in terminal or by coding from scratch as instructed above.
   

*** Testing out the Model using Webcam ***
6. Now, that we've created our classification model we can directly apply it to live-feed video using opencv.
7. Desgin Bounding Box, Texts and etc if needed.
8. Run python detectMaskLive to start the webcam and check whether You're Wearing a Mask or Not!. Running the command shows up something like this -

***** Loading FaceDetector *****
***** Loading FaceMask Detector *****

------------- Starting Video Stream -------------
------------- Video Stream Stopped -------------

Next Updates

To Do
1. Improve Accuracy
2. Improve Latency Time

NOTE : It's noted that increasing the frame window aspect ratio decreases the model accuracy.

Contributing

If you are the helping and contributing one, your efforts and suggestion are always welcomed.

Reference(s)

1. PyImageSearch
2. Datasets