| With the rapid development of deep learning technology,the neural network applied to face recognition technology is flourishing.The accuracy of most face recognition within the controllable range has reached very ideal results.However,face images acquired under real conditions are often difficult to recognize because of light occlusion,physical occlusion,and self-occlusion.The new crown epidemic that broke out at the end of 2019 has forced everyone into the stage of wearing masks.This paper focuses on the problem of low accuracy and poor stability of face recognition under occlusion conditions.Specifically:(1)Add a lightweight channel attention mechanism ECA module after the convolutional layer in the original RepVGG network to improve the recognition accuracy of the network model;(2)Use Softpool in the pooling layer to replace the original pooling method to improve accuracy;(3)Replace the traditional cross-entropy loss function with the Focal loss function to enhance the learning performance of the network.In order to verify the effectiveness of the improved algorithm in this paper,three experiments were carried out on the AR face dataset:(1)Face key point detection experiment based on Haar classifier;(2)Based on the contrast experiment of occluded face recognition under the RepVGG network model before and after the improvement,the ablation experiment with different improvement points,and the ten-fold cross-validation experiment after the improved model;(3)The occlusion detection experiment of the occluded face based on the YOLOv5 network model and the face recognition experiment with different occlusions under the improved RepVGG network model.The above experimental results show that the improved algorithm has a relatively high recognition rate for the common types of occlusions on the eyes and mouths of the face,as well as for occlusions caused by changes in illumination and expression,and improves the recognition accuracy and speed of the network model. |
PDF Full Text Request