Research Of Mask Wearing Detection Algorithm Based On Improved YOLO

Since the outbreak of the COVID-19 epidemic,people’s awareness of health and safety has gradually increased.Wearing masks in public places is a universal means of protection.In response to the demand for intelligent detection of facial mask wearing in some public places,this paper applies the algorithms of YOLO series based on deep learning to the task of mask wearing detection for research and improvement.In order to improve the detection performance of existing mask wearing detection algorithms,especially for non-standard wearing categories,the anchor-based YOLOv3 algorithm and the anchor-free YOLOX-tiny algorithm are improved to enhance the performance of the model in locating and classifying mask wearing detection tasks in input images.The main research contents and results are as follows:(1)Aiming at the problems of slow detection speed and many network parameters of YOLOv3 algorithm,this paper proposes a lightweight mask wearing detection algorithm,which improves detection performance by optimizing the original backbone,neck and detector.A lightweight feature extraction network is introduced to speed up feature extraction,a multi-scale attention mechanism is added to enhance feature extraction capabilities.Aiming at the problems of the misalignment between deep and shallow feature information in the feature fusion stage,a feature alignment and selection module is designed to improve the network’s detection rate of faces and masks.The feature enhancement module is proposed to strengthen the attention on the arrangement information of the facial and mask positions.The decoupling channels are constructed to improve the recognition accuracy and convergence speed of the improved algorithm.The experimental results on the mask wearing detection dataset show that the improved algorithm achieves a detection accuracy of 93.2%,which is 2.5% and 5.6% higher than YOLOv3 and lightweight algorithm YOLOv3-tiny,respectively.The detection speed reaches 107 FPS,and the network parameter quantity is reduced by 52.8% compared to YOLOv3-tiny(2)Aiming at the problem of feature information loss caused by the use of traditional spatial pyramid pooling structure in the feature extraction stage of the YOLOX-tiny algorithm,SF-CSP is designed to replace the SPP structure,effectively improving detection accuracy while accelerating the downsampling process.SIOU positioning loss function is introduced to increase the angle constraint and shape constraint of prediction box regression in training,which makes the network converge faster in training and can also promote the improvement of accuracy.Vari Focalloss is introduced to alleviate the problem of imbalanced positive and negative samples in the dataset,which can enhance loss optimization for high-quality samples,and thus improve detection accuracy.The detection accuracy of the improved algorithm has increased by 2.3 percentage points compared to the original algorithm in the mask wearing detection dataset,reaching 92.8%.The improved YOLOX-tiny algorithm achieved a detection accuracy of 80.7% on public datasets,effectively verifying its reliable generalization and universality.