Improving Performance of Yolov5n v6.0 for Face Mask Detection

• Published in: 2024 International Conference on Artificial Intelligence in Information and Communication (ICAIIC) pp. 129 - 134
• Main Authors: Al-Shamdeen, Muna Jaffer, Ramo, Fawziya Mahmood
• Format: Conference Proceeding
• Language: English
• Published: IEEE 19-02-2024
• Subjects: COVID-19; Deep Learning; Feature extraction; Measurement; Object Detection; Pandemics; Predictive models; Proposals; Resnet; Security; YOLOv5; YOLOv5n v6.0
• ISSN: 2831-6983
• DOI: 10.1109/ICAIIC60209.2024.10463515

The COVID-19 coronavirus pandemic has generated a global health crisis in all Worldwide. According to the World Health Organization (WHO), protection against COVID-19 infection is an essential countermeasure. one of the most effective countermeasures is wearing a facial mask which is imperative in our everyday activities, particularly in communal settings, to mitigate the transmission of the illness. In this study, we have enhanced the architectural design of YOLOv5n v6.0 for face mask detection by constructing a modified model known as Proposal YOLOv5n v6.0 model. The primary objective of this modification is to enhance the feature extraction and prediction capabilities of the YOLOv5n v6.0 model. In our proposal, we outline the integration of a residual network (ResNet) backbone into the YOLOv5n v6.0 architecture by replacing the first three layer of YOLOv5n v6.0with ResNet_Stem module and ResNet_Block module to enhance the feature extraction capabilities of the model and replace Spatial Pyramid Pooling Fast (SPPF) module in original model with Spatial Pyramid Pooling-Cross Stage Partial (SPPCSP) modules which combines SPP and CSP to create a network that is both effective and efficient. In our proposal model we have carefully curated a set of anchor configurations tailored to the specific requirements of small object mask detection. MJFR dataset was used for testing and evaluation of the original and proposed model which consist of 23,621 images and collected by the authors of this paper. The performance of both models is evaluated using the following metrics: mean average precision (mAP50), mAP50-95, recall (R) and precision (P). We conclude that the proposed model outperforms the original model in terms of accuracy for mAP50, mAP50-95 which is the metric used to measure the performance of the object detection model.