Face Mask Detection

Arpit Chauhan

Published in

Analytics Vidhya

4 min readNov 17, 2020

Face Mask Detection is a project based on Artificial Intelligence . In this we detect people with or without mask .

In making of this project we have two phases .

Train model using Convolution or any pretrained model which detect face masks in images .
Then, Detect faces in video or images and get prediction from our trained model .

In this Project I used Transfer learning to train model on dataset in which I used pretrained model MobileNetv2 , Now question is what MobileNetV2 is ? what is the Strucutre of it ?

MobileNetV2

MobileNets are small, low-latency, low-power models parameterized to meet the resource constraints of a variety of use cases. They can be built upon for classification, detection, embeddings and segmentation similar to how other popular large scale models, MobileNetv2 is a Significant Improvement over MobileNetV1 . The Structure of MobileNetV1 and V2 is below .

Now , Come to Code and Data Explanation

Dataset

In the Dataset we have two types of images one who wore mask and other which do not have mask.

Training Dataset consist of 4296 images in which 2455 images have label with mask and 1841 images are of without mask faces and Validation Dataset consist of 300 .

Import :

First I import Required modules like tensorflow , keras ,optimizer , layers and pretrained model(MobileNetV2)

Loading Data And Augment Image :

Now I load all images for training and preprocess using MobileNetv2 “preprocess_input” through which the images are ready as mobilenetv2 require .

we also Augment our images through which we can get more data and variety in data. we add augmentation technique like zoom , horizontal and vertical flip , rotation .

After this step we have training and validation batches of shapes which mobilenetv2 require as input .

Loading Pretrained Model :

As I previously write I Used MobileNetV2 . So , I download weights of model and create a object of MobileNetV2 type after we freeze the layers of our pretrained model through which layer weights can’t be modified when our model is in training .

Adding Some More layers To Model :

Adding Some Layer at the end of the model to achieve good accuracy or save model from overfitting .

At last add one fully connected layer which contains neurons equal to number of labels we have (In this case we have 2 labels mask or no mask) . And After Define Model Input and output layer .

Now Our Model Structure is ready .

Compile and Train :

Now it’s time to train the model but before training we have to define loss function and optimizer i.e., compiling a model .

After Hypertuning of parameters I found good results with Learning Rate 0.00001 . So , I used Adam Optimizer with learning Rate 0.00001 and loss function binary_crossentropy . then after I trained the model with 15 epochs and validate it on validation data (which contains 300 images) .

Testing Phase :

Now we have to test model on test data through which we got good idea weather model will work good on real time data or not . so , we have 74 images of 2 classes as test dataset after evaluating it on test dataset I got good results .

Loss : 0.058

Confusion Matrix :

A confusion matrix is a table that is often used to describe the performance of a classification model on a set of test data for which the true values are known.

After Plotting Confusion Matrix what I see My Trained Model only predicts 1 image wrong else all images are correctly predicted . so model will work Good in Realtime .