Implementation of Face Mask Detector with OpenCV, Keras/TensorFlow, and Deep Learning
The purpose of this system is to implement custom deep learning model to detect person with mask, or person with no mask.
Typical application is to implement on mobile for personal safety (static images), and monitor crowds during COVID-19 pandemic (video stream).
Phase 1
- Load face mask dataset
- Train face mask classifier with Keras/Tensorflow
- Serialize face mask classifier to disk
Phase 2
- Load face mask classifier from disk
- Detect faces in image/video stream
- Extract each face region of interest (ROI)
- Apply classifier to each face ROI to determine “mask” or “no mask”
- Show results
Publicly available images dataset is used to train the model. It contain 1,376 images belonging to two classes: 690 images with_mask and 686 images without_mask.
It is pertinent to note that person with masks are different from person without mask, otherwise model becomes heavily biased and fails to generalize.
- Take image of person without mask
- Find facial landmarks (eyes, eyebrows, nose, mouth, jawline)
- Apply mask to facial landmark ROI
Images without Mask
Images with Mask
Get image of person without mask
Compute bounding box location of face
Face ROI
Apply landmarks to localize eyes, nose, mouth, etc
Overlay mask on detected facial landmarks
Image of person with mask
MobileNetV2 is a general architecture
- can use different input layer size and different width factors
- reduces inference cost on mobile devices
- highly efficient on embedded devices with low computation
Similar to the original MobileNet
- lower parameter count than the original MobileNet
- supports any input size greater than 32 x 32
- larger image sizes offer better performance
Uses tensorflow.keras
- Date augmentation (rotate, zoom, height/width shift, shear, horizontal flip)
- Load MobileNetV2 classifier (with fine-tuned ImageNet weights)
- Build fully-connected head
- Pre-processing (resize image to 224×224 pixels, conversion to array format, scaling the pixel intensities in the input image to the range [-1, 1])
- Loading image data
Uses sklearn
- Apply binary class labels
- Split dataset (80% training data, 20% test data, shuffle data before splitting)
- Print classification report
Uses imutils to find and list image in dataset
Uses matplotlib to plot training loss and accuracy curves
- Initial learning rate: 10-5
- Epochs: 20
- Batch size: 32
Image input: 224x224 (3-channel)
Default fully connected layer of MobileNetV2 is not used
Custom implementation of fully connected layers
- Pool size: 7x7
- Dense layer: 128 units, activation: ReLU
- Dropout rate: 0.5
- Dense layer: 2 units, activation: Softmax
Training parameters
- Loss function: binary cross entropy (binary output i.e. “mask” or “no mask”)
- Optimizer: Adam (computationally efficient, little memory requirement)
- Metrics: accuracy
First, make sure required modules are install by running, then follow 3-steps given below to train and predict mask in static image or video stream.
pip install -r requirements.txtStep 1: Train model
python train_mask_detector.py --dataset datasetStep 2: Detect face mask in static image
python detect_mask_image.py --image examples/example_01.pngStep 3: Detect face mask in video stream
python detect_mask_video.py| Infrastructure | Specification |
|---|---|
| Operation System | Windows 10 Pro (64-bit) |
| Processor | Intel® Core™ i7-5500 CPU @ 2.4 GHz |
| RAM | 16 GB |
| GPU | GeForce 830M (2 GB), Compute Capability 5.0 |
| CUDA | Toolkit version 10.1.243, Driver version 446.14 |
| Loss | Accuracy | |
|---|---|---|
| Training | 0.1037 | 0.9747 |
| Validation | 0.0656 | 0.9891 |
| Limitation | Reason | Way Forward |
|---|---|---|
| Size of dataset | 1,376 images is not enough to apply this train model in production. | Larger dataset is required and retraining of model will overcome this limitation. |
| Face mask is not detected | Facial landmark ROI is basis of face detection, if face is mostly hidden behind mask then model may show false negative. | Diverse images with mask are required to overcome this limitation. |
| Artificially created images with mask | Public dataset of images with mask is not available. | Create a dataset containing real images with masks, and retrain model. |
Copyright 2020 Hamza Rashid
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.