Face Recognition Based On The Combination Of Deep Learning And Traditional Methods

In recent years,with the improvement of the level of biomedicine and computer technology,face recognition technology has attracted wide attention from researchers.As a type of biometric technology,it has the characteristics of being natural and not easily perceived by the individual being tested.It has become a research hotspot in the field of computer vision and is widely used in security inspection,mobile payment and other fields.Face images are easily affected by factors such as illumination,expression,occlusion,and posture changes.These uncertain factors greatly increase the difficulty of face recognition.Since Gabor transform has strong robustness when extracting local features of face,it is often used for classification and recognition of extracting local features of face.Compared with traditional face recognition,deep learning method can better extract the deep global features of face images without manual involvement.Based on these two methods,this paper proposed a face recognition method combining local features and global features.In this paper,Gabor wavelet was used as a method for extracting local features of traditional face images,and the concept of image segmentation was introduced,which was combined with two-dimensional Gabor wavelet to obtain local features.Since the local feature dimension after Gabor transformation was too large,the features of key parts of the face(such as eyes,nose,mouth,etc.)were grouped and feature fusion was used to give different face parts different weights.The key facial features were fused,so that the extracted high-dimensional features were reduced to low-dimensional features to recognize the face.In terms of deep learning,the deep residual network was selected as the basic structure for extracting global features.On the basis of the original residual network,the size of the network model was determined by modifying the number of residual units and increasing the network width,and training on the data set for parameter tuning and selection,selecting the appropriate parameters,The network model was also optimized for key network settings to extract more effective depth features.In order to verify the performance improvement of the improved model S-ResNet-14 compared to the original ResNet-32 network model,an experimental comparison was performed on the data set Extend Yale B.The results showed that when selecting the data set 80% of the face when the image was used as the training set and the remaining samples were used for testing,the improved S-ResNet-14 network achieved an accuracy rate of 94.4%.Compared with the original ResNet-32 network,the accuracy rate was improved by 0.17%.Finally,based on the idea of feature fusion,the local features extracted by traditional methods and the global features extracted by the residual network were fused.By conducting a classification and recognition test on the public face database Extend Yale B,an optimal recognition rate of 99.15% was achieved.The experimental results showed that the improved algorithm proposed in this paper can obtain a higher recognition rate under the influence of facial expressions such as expression,posture,occlusion and other factors.Figure [45] Table [8] Reference [51]...