Research And Application Of Face Occlusion Recognition Algorithm

This thesis focuses on fast,efficient,and accurate mask wearing detection in places with huge traffic flow,such as airports,railway stations,and supermarkets,during the period of epidemic prevention and control.Based on innovative applications of pattern recognition,deep learning,and machine learning in epidemic prevention and detection work,the FRF-MR algorithm and the YOLOv5-FOD algorithm are proposed to address the questions of whether people wear masks correctly and whether they use other objects to cover their faces in order to fool machine detection.The main work includes:（1）This thesis integrates the open source dataset from Wuhan University,Wide Face dataset from Chinese University of Hong Kong,MAFA dataset from Institute of Information Engineering,Chinese Academy of Sciences,and image dataset from web platform for whether the mask is worn correctly,and the data are manually annotated according to PASCAL VOC standard data format using Labelimg software to generate XML files for training.In this thesis,we first preprocess the data to improve the speed and accuracy of training,and introduce the datasets to extend the data.In this thesis,the experimental data are horizontally flipped,image scaled,brightness boosted,saturation boosted and contrast boosted to improve the dataset and improve the network robustness.（2）To address the problem that the existing algorithms do not consider the influence of different image specifications in the data set,which leads to low detection accuracy and slow detection speed,this thesis proposes a face mask recognition algorithm FRF-MR,which uses feature pyramids for fast target detection and has the advantages of fast and high accuracy for multi-scale small target detection,and this algorithm solves the problems of slow detection speed and low accuracy due to factors such as variable target areas and different image specifications in the face mask detection problem.The algorithm solves the problems of slow detection speed and low accuracy due to factors such as irregular target area and different image specifications in face mask detection problem.Experiments prove that the algorithm achieves 96% detection accuracy and 86% recall rate in the dataset.（3）Considering the problem of multi-label detection with multiple classifications in the dataset,a face mask target detection algorithm YOLOv5-FOD is proposed,which has the performance of fast and accurate detection of multi-label classification targets.Since in face mask recognition,it is not only necessary to recognize the face target and accurately distinguish whether the target is wearing the mask correctly or not,but also to accurately distinguish the authenticity of the target wearing the mask.The algorithm performs fast and accurate detection of the authenticity of face masks and solves the problems of low accuracy and slow detection speed of multi-label class face masks.The experiment proves that the algorithm will achieve fast and efficient detection of multi-label classification face masks.