Introduction :

In today’s world, face recognition has become an integral part of our lives. We use face recognition feature for multi-purposes in our latest gadgets such as FaceID, etc. In this project, we will implement face recognition technique using deep learning concepts. Our aim is to recognize people from the given dataset such as photos or videos. The name of that person will be displayed on the photo or video being displayed. We will accomplish this through four important phases: face detection, posing and projecting faces, encoding (recognizing) faces through basic facial measurements and finding person’s name from encoding. We will use deep convolutional neural networks for encoding faces and also linear support vector machine (SVM) classifier for getting the information about the name of the person and displaying it. We also want to map a person’s face, even when the whole face (50%) is not displayed in the dataset. Thereby, predicting a face and identifying it. In this way, we can identify a person when he/she has worn a face mask, especially in current times of COVID-19 pandemic. We can use this feature in many gadgets and applications for authentication and medical purposes.

Expected Results:

The results expected from each stages is:

1^st stage: We are able to recognize faces from an image or video through mapping of HOG patterns.

2^nd stage: No matter how the faces are posed or projected, we are able to locate eyes and mouth of a face properly, which makes recognizing faces more accurate.

3^rd stage: In this, We do the mapping of the faces which are incomplete. We predict the whole face of a person by using half of his/her face's image as data. Once the mapping is done, we then, use it for encoding. We encode the faces, and are able to get specific unique measurements of a person’s face, which distinguishes a face from the other.

4^th stage: In the last stage, we get our output of the project. We are able to display the name of the face being detected in the image or video after running the classifier. Hence, able to recognize the faces of different people from our database.

The coronavirus was declared a pandemic by WHO on March 11th, 2020, and by that time, we already had 118,000 confirmed cases and had spread to 114 countries around the globe. Since then, due to a variety of reasons some of which include strict measures not being put into place and government negligence, we have seen these figures blown out of proportions. Currently, at the time of writing this, we have 16 million confirmed cases and near 645,000 deaths worldwide. While some factors can be eluded as being not being under our control as individuals, we cannot say that face masks do not have a big role to play in controlling this pandemic.

Data Collection and Preprocessing

1. Data Collection for Training

The total number of images in the training set were divided into the two categories as followe:

1. with mask : 1938 images

2. without mask: 1930 images

Model Fitting

Batch size: 32

Epochs: 2

Learning rate : 1e-4

Gradient Descent Optimizer : Adam

Loss function : Sparse cross entropy

Metric evolution : F1-score

We checked all photos that include in our dataset that we get from online and we used some kaggle dataset include crowd based settiong of people. Applied RetinaNet face model for face detection to generate detected face crops from the input image.

Results

Results obtained on the classification model:

Loss - 0.0182

F1 score: 0.99