Research And Application Of Face Recognition Technology Based On Deep Learning

Face recognition,as one of the research hotspots in the field of computer vision,is an intelligent authentication technology that simulates computer to extract and identify facial features.Compared with fingerprint,iris and other biometric recognition technologies,face recognition is widely used in criminal detection,security,human-computer interaction and other fields for its convenience,reliability,non-contact and other advantages.With the rise of deep learning,face recognition technology has been greatly developed,especially the face recognition method based on deep convolution neural network has become one of the mainstream research methods.However,in practical application,the performance of current face recognition still needs to be further improved due to the influence of expression,posture,illumination and other internal variable factors and external interference factors.Therefore,this paper focuses on the deep convolution neural network and further improvements are made in view of the shortcomings of the existing face recognition methods so that the recognition accuracy can be improved.The main work includes:(1)Aiming at the two shortcomings of traditional pooling methods,that is,the loss of spatial information and the equal treatment of each element in the local pooling area of input feature map,this paper improves the traditional pooling method by using the strategy of local concern,and designs a local weighted average pooling method.By assigning different adaptive contribution weights to different elements in the locally pooled region,the method achieves the purpose of effectively compressing the feature map and reducing the information loss.And then the improved pooling method is applied to FaceNet for face recognition.The experimental results show that recognition accuracy of FaceNet Face recognition algorithms based on local weighted average pooling is improved by nearly 1% on the original basis,and it has good robustness to the changes of expression,illumination and posture.(2)In view of the fact that the convolutional layer has the disadvantage of treating different regions of the image when extracting facial image features,this paper uses the global attention strategy to improve the traditional convolution operation,and designs a global dynamic convolution operation method.The method expresses the difference in the role of each image region in facial recognition by assigning different learnable weights to different regions of the face image,thereby extracting more discriminative facial features.And then the improved convolution operation method is applied to face recognition in the initial convolution layer of FaceNet.The experimental results show that the improved FaceNet face recognition algorithm based on global dynamic convolution can achieve a small improvement in accuracy without increasing the parameters.(3)Combining the improved local weighted average pooling with the global dynamic convolution operation method,an improved FaceNet face recognition algorithm based on local and global attention is designed.Compared with the above two improved algorithms,the improved FaceNet face recognition algorithm based on local and global attention can obtain better recognition accuracy on the same data set.(4)A real-time face recognition system consisting of three modules,image acquisition,face detection and face recognition,is designed by using the improved algorithm model based on local and global concerns,which further verifies the effectiveness and practicability of the algorithm.