One to many faces detection on mask-wearing individuals

Abstract

In this project, we were interested in creating a system that allows to identify people wearing masks. This particularity will allow corporate companies and enterprises to have a solution to the current problem Covid-19 has presented us with. Firstly, we started by studying the current history of facial detection and what it revolves around,we’ve made sure to mention the different types of facial detection method an shed some light on the advantages of each one. Then a deep dive was made in the theoretic notions used in the facial detection of our individuals , where an introduction to neural networks and machine learning has been made. Secondly, a reduction of our research radius has been made, where a focus on mask-wearing individuals was pointed out. This latter included notions such as AI frameworks where we’ve compared each of Custom vision as well as Tensorflow. Talking about these frameworks allowed us to mention the graphic libraries that are availble in this day and age such as OpenCV, where a study was carried out to further research the prediction confidence of each library this step allowed us to have a better understanding on the need of filtrating our dataset. Once our dataset established we’ve concluded from the various detection classifiers (MobileNetV2 being a solid candidate) and optimizers that are available to us that using CNN is the way to go for this system. Which brings us to the third chapter of this research, where we’ve managed to divide our model into two phases, the first one includes loading our dataset into the system whilst the second one, is about loading the face mask classifier model which will allow us to apply our chosen classifier. Our system’s procedure is then layed-out in four different distinct steps: acquiring our dataset, learning and modeling it, developing and training our model then finally deploying our model for live feeds such as webcams. It seems that using the CNN architecture along functions from the Keras machine learning library, is a an important step since it allows for faster results when tuned using MobileNetV2. Then, on the fourth chapter we went on the implementation phase where we applied the theoric notions and methods explored in past chapters, along with attempts and trials made with a variety of ML techniques. This has led us to conclude on a model generated with an addition of the MobileNetV2 classifier and two Conv2D layers that gave us the better results. The accuracy of the neural network on image datasets is almost always over 97 percent, which can be improved in the future if we decide to deprecate MobileNetV2 or not. The model we generated can be used by various government agencies in crowded areas such as markets, airports, railway stations, and other crowded areas to ensure that people follow safety precautions. This model could be embedded with an alarm or buzzer in the future, and IoT could be implemented using deep learning CNN. When a person is detected not wearing a mask, the device will sound an alarm. If this model is implemented correctly, it will assist in ensuring human safety in this global pandemic.