Research On Face Recognition Method Based On Multiscale Feature Dimension And Manifold Learning

Face recognition technology,as a convenient,reliable,and secure biometric recognition technology,has a broad application market in security monitoring,access control recognition,and public safety.In practical applications,facial images undergo numerous changes due to various factors,posing a huge challenge to facial recognition.This article will use convolutional neural networks and manifold learning as the framework,mainly to solve the problems of high exposure,blurring,occlusion,low-quality images,low memory,low computing power,and high real-time requirements in face recognition.The main research content is as follows:(1)This thesis proposes a convolutional neural network model based on Ineception-V3 and Transformer models for face recognition of high exposure,blurring,occlusion,and lowquality images.By using convolutional kernels of different sizes and asymmetries,such as1×1 、 3 ×3 、 5 ×5 、 3 ×1 、 1×3 and so on,multi-scale features are obtained and fused.Firstly,in order to improve the quality of facial input images and enhance the performance of the network model,image preprocessing operations are carried out on the input facial images;Secondly,the convolution kernel of the intermediate feature extraction layer of Inception V3 model is improved to widen the network,and multi type convolution kernels are used to obtain multi-scale facial image features,and Concat is used to connect and fuse them to obtain more abundant Semantic information of facial features;At the same time,Transformer was used to replace the fully connected layer,integrating the overall and local information of facial images,and reducing the number of model parameters;Finally,through comparative experiments with a single network model,it can be seen that the improved network model has improved facial recognition rate,accuracy,and speed.(2)In response to the large number of model parameters proposed in the previous section,as well as the current low memory,low computational power,and high real-time requirements of terminal devices,it is particularly important to efficiently mine low dimensional manifold spaces for facial features.This thesis proposes a lightweight neural network model incorporating manifold structures.Firstly,this article directly adopts a lightweight neural network model,which uses a parallel multi convolutional kernel deep separable convolution to extract facial features.This method can reduce computational complexity and also introduce more nonlinear structures;Secondly,in order to extend the Euclidean space to the Non Euclidean space,the data format needs to be changed,so the covariance of the extracted face features is calculated,and some regularization processing is carried out to make the face features located in the Riemannian manifold space;Finally,the Riemannian manifold structure of the symmetric positive definite matrix is integrated into the lightweight neural network.Experiments show that the network model can reduce the number of parameters of the model,achieve real-time effect in running speed,and improve the accuracy of recognition.