Research On Generation And Recognition Method Of Disguised Face

Face recognition refers to the use of computer vision and image processing related technologies to extract different features of the face for matching and recognition using face images as a medium.It is currently one of the most popular research topics in the field of computer vision.At the same time,as an inherent attribute of people,human faces are one of the ideal choices for people’s identity verification.Face recognition technology has been widely used in fields such as national security,smart access control,and public security deployment control.At present,the recognition accuracy of face recognition algorithms based on deep learning has surpassed the recognition accuracy of human eyes on ordinary face data sets.However,when the face image has intentional or unintentional disguised occlusion,recognition will become difficult,and the recognition accuracy of the algorithm will decrease.Moreover,due to the fact that there are fewer publicly disguised face data sets,the performance of the training model is poor,which is one of the important problems in the field of disguised face recognition.This paper studies the above-mentioned problems and proposes a method to generate a disguised face image,and improves on the FaceNet network architecture and proposes a disguised face recognition algorithm.The specific work and contributions are as follows:(1)This article summarizes the traditional face recognition methods and the face recognition methods based on deep learning,and explains the difficulties of disguised face recognition.At the same time,the specific process of face recognition and commonly used face recognition data sets are introduced.(2)Aiming at the problem that there are few disguised face datasets that are currently publicly available,this paper proposes a method for generating disguised face images based on a generative confrontation network,using the Celeb A dataset as the original dataset.This method can edit the face attributes of the face images in the data set,and generate a disguised face image by adding sunglasses and a beard,so as to form a disguised face data set together with the original image.The expansion of the disguised face data set improves the over-fitting problem of the model and improves the generalization ability and performance of the recognition model.(3)This paper proposes a lightweight disguised face recognition network model based on FaceNet.First,the deep neural network part of the FaceNet network architecture is improved,and a deep separable convolution module is added to the backbone feature extraction network.The network reduces the amount of model parameters through deep separable convolution,which improves the speed of model calculations.Next,this article introduces the attention mechanism into the lightweight network model constructed.The attention mechanism is implemented by adding a CBAM module.CBAM calculates the channel and space dimensions of the input feature map,calculates the weights of different regions in the feature map,and then distinguishes the importance of different regions according to the weights.The attention mechanism can strengthen the features of the visible area in the disguised face image,ignore some irrelevant features,and effectively improve the accuracy of the disguised face recognition algorithm.Finally,this article uses the cross-entropy loss function and the Triplet loss function as the overall loss function to construct a classifier to assist the network to converge,which effectively improves the convergence of the network.