Mask detection using deep learning method

• Main Author: Tan, Yu Xuan
• Format: Thesis
• Language: English
• Published: 2022
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://eprints.utm.my/id/eprint/99576/1/TanYuXuanMSKE2022.pdf; http://eprints.utm.my/id/eprint/99576/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:149778

Wearing face masks outdoors has been a new norm due to the COVID 19 pandemic as an initiative of controlling the spread of coronavirus. To reduce the risk of people being exposed to viruses, face masks were compulsory to be worn by Malaysians. However, there are people who refused to do so due to various reasons such as feeling lazy, uncomfortable, troublesome, and others, even the act of wearing a face mask is enforced by law. Therefore, it is essential to build a face mask detector to monitor automatically and ensure people are wearing masks correctly. The performance such as precision and response time of face mask detectors are important to support their application in the real-time working environment. The issue of performance enhancement in the form of adding more layers or implementing hybrid models such as spatial pyramid pooling (SPP) modules is increasing the complexity of the algorithm and making it bulky. The objective of this paper is to build a face mask detector by using the latest high-performance deep learning model, YOLOv4 and YOLOv5 together with MixUp technique which can contribute to high mean accuracy precision (mAP) and short inference time that suffice the requirements to be working in a real-time environment. This research conducted data sets collection and data annotation at the beginning stage of the algorithm, then MixUp technique was applied to the collected datasets to train the YOLOv4 and YOLOv5 using Google Colab. Next, the trained model was tested, and the performance was evaluated in terms of mAP using the average precision (AP) from the confusion matrix and inference time based on the time taken for prediction. The algorithm with the YOLOv5 model having slightly lower mAP than YOLOv4 but shorter training and inference time. However, both models able to detect and classify the input image to three classes included with-mask (1), without-mask (2), and incorrectly with-mask (3) with good performance.