Research On Key Technologies Of Face Recognition Under Mask Occlusion Based On Deep Learning

The COVID-19 virus outbreak swept the world in December 2019.Under the current situation of epidemic prevention and control,traditional face-recognition technology has been challenged by the widespread use of face masks when people travel.In this context,this paper,the face recognition technology of wearing masks based on deep learning is studied,and an improved algorithm is proposed.Experimental results show that our algorithm has better performance than the traditional algorithm.Specific research contents are as follows:(1)A mask-wearing face detection technique based on deep learning is improved.technology.The R-Net and O-Net in MTCNN are analyzed,and NMS algorithm in its network is introduced into the soft-NMS algorithm and Io U-aware technology to obtain the mask-NMS optimization algorithm,the algorithm improvement It is mainly to define the confidence in NMS to select the accumulation of the classification probability score and the Io U,the optimized algorithm makes it not directly canceling the candidate regression box that has exceeded the threshold,but reduces the candidate regression box of confidence level.The experimental results demonstrate the effectiveness of our algorithm our algorithm has dropped by 0.79%,and the application is increased by 2.07%on the recognition accuracy of the face recognition model.(2)A mask-wearing face image quality assessment technique based on deep learning is improved.Collecting data and utilizing data enhanced algorithms constructs different quality wear mask face image data sets are 10866 and 14565 respectively.An 8-layer CNN model is built,and the output is the confidence level of high-quality mask wearing face image,which is taken as the image quality score.The end of the training experiment is high,and the precision of the classification of lower quality wear masks is97.93% and 97.67%,and the average accuracy is 97.78%.(3)A mask-wearing face recognition technology based on deep learning is improved.The structure of VGG-16 Net and various convolution layers are studied,and introduced an algorithm based on the weight of the receptive field,which a 2-dimensional Gaussian nuclear function is designed according to the filter characteristics of the 2-dimensional Gaussian nuclear function.Setting parameters (d= 7,?= 3 can make the shadow area weight in the field decrease,and the weight of face effective feature information is increased.Its application is on the face recognition model,and the experimental demonstrate that our method recognition accuracy increased by 5.81%.