无监督数据集子类划分的人脸口罩佩戴识别算法

人脸口罩佩戴识别成为疫情防控的一项重要手段，而目前口罩佩戴检测主要还是通过人工监测，基于深度学习的口罩佩戴检测系统较少，且存在误检、漏检和检测速度慢等问题。针对口罩佩戴检测中不规范佩戴口罩数据集较少，和对检测精度和检测速度要求较高的实际应用需求，从数据集和网络两方面改进人脸口罩佩戴检测方法:通过在无监督自分类方法中引入标签矫正算法对数据集进行子类划分，减少数据集类内差异，提高网络检测精度;调整目标检测网络结构，去除小尺度检测的网络层，提高网络检测速度;引入注意力机制模块，增强网络对细节特征的提取能力，提高网络检测精度。口罩佩戴情况的平均检测精度从79.34%提升到93.12%，检测速度提高了6.4%，设计的网络结构能够满足实际应用的需求。

Face mask wearing recognition algorithm based on subclass division of unsupervised data set

Face mask-wearing recognition has become important means of epidemic prevention and control. However, mask-wearing detection is mainly conducted through manual monitoring, and at present, there are few high-precision and real-time mask-wearing detection systems based on deep learning, which is still with the problems of false detection, miss detection, and slow detection speed. According to the fact that there are few data sets for wearing mask recognition as well as the requirements of high detection accuracy and detection speed in mask-wearing recognition, the face mask-wearing detection method is improved from two aspects of data sets optimization and network design. Firstly, the label correction algorithm is introduced by the unsupervised self-classification method, the subclass of the mask-wearing data set is automatically obtained, and the detection accuracy is improved. Secondly, by adjusting the structure of the target detection network, the network layer of small-scale detection is removed to accelerate the network detection speed. Moreover, in order to improve the classification ability of data categories with the small differences among classes, the attention mechanism is introduced to enhance the extraction of detail features. The detection results show that the average detection accuracy of mask-wearing is improved from 79.34% to 93.12%, and the detection speed is improved by 6.4%. In conclusion, the designed network can satisfy the demand of practical application.