REAL-TIME FACE MASK DETECTION USING CASCADED BI-LEVEL FEATURE EXTRACTION TECHNIQUE FOR MANAGING THE SPREAD OF COVID-19

Abstract

COVID-19 is a respiratory sickness that dealt the human world with one of the deadliest blow in 2020, it is one of the pandemics that has threaten the very existence of humanity in recent times. Its fast spread has caused widespread devastation and infected tens of millions of people around the world. Due to the lack of specific cure for COVID-19, wearing a face mask has proven to be an effective method of reducing its transmission. This is now required in most public venues, resulting in an increase in the demand for automatic real-time mask detection devices to replace manual reminders. This is because people are not willing to wear the face mask and those who do, are not likely to do so the recommended way. This lead scientists and researchers to integrate surveillance technology with Artificial Intelligence (AI) to create a system that identifies people wearing face mask in public areas. Face mask detection necessitates a large amount of data to be processed in real-time or on devices with limited processing resources, therefore local descriptors that are fast to calculate, fast to match, and memory efficient are in high demand. The goal of this research work is to reduce the continuous spread of the deadly pandemic by creating a face mask identification model to classify face images into face mask present and face mask absent. Therefore, this study offers a cascade of Features from Accelerated Segment Test (FAST) corner detector and Histogram of Oriented Gradient (HOG) feature descriptor to allow faster matching and minimize memory usage and computation cost. To achieve this, the images were preprocessed by performing facial landmark identification using viola Jones algorithm. The resultant images were converted to grey scale images and finally, the images were smoothed by the application of median filter also known as median blur. This filter removes noise from the images. After all the preprocessing steps were carefully carried out, the images were passed to the FAST corner detector to detect the point of interests. These interest points were then passed as input to HOG for feature description. The features were then classified into face mask present and face mask absent using Support Vector Machine (SVM), Naïve Bayes (NB) and Convolutional Neural Network (CNN). The results obtained had a 99.41% accuracy, which was higher than the prior work's 99.27% and 95% accuracy. In addition, the suggested method extracted face features in 48 seconds for training and testing. This study demonstrated that the technique is capable of detecting face masks in real-time.