| As a research hotspot in the field of computer vision,face recognition technology has ushered in technological innovation along with the development of artificial intelligence.Due to the excellent performance of deep learning in image processing,breakthroughs in face recognition technology have been made,especially in the face recognition technology based on deep convolutional neural networks,which has made remarkable achievements.Traditional face recognition methods require artificially set expression features and complex timeconsuming feature extraction,while methods based on deep convolutional neural networks only need to design an efficient network model and train on a large number of data sets.By automatically learning to extract high-dimensional invisible features,you can obtain excellent classification accuracy.The performance and effect of this method mainly depend on the design of the network structure.Therefore,the research focus of this paper is to solve the problem of deep convolutional neural network face recognition algorithm with complex network structure,huge amount of parameters,and difficult training.The network structure also provides effective training methods to make it quickly converge and stabilize and improve the recognition accuracy.The main work and innovations are as follows:(1)For the problem of large parameters of the VGG-16deep convolutional neural network model,in order to reduce the model parameters,this paper designs and implements a VGG-16 network structure that optimizes the fully connected layer,reducing 86%of its total parameter amount The first two fully connected layers have increased the number of convolution kernels in Group5 and adjusted the underlying structure of the network.The improved VGG-16 network has fewer layers and can effectively extract deep invisible features in more dimensions.Under the premise of ensuring the face recognition rate,the overall network parameters are reasonably reduced.(2)For the problem that the improved VGG-16 network does not match the original training method,in order to ensure the best performance performance,this paper gives a retraining method.The method adjusts adaptively from the aspects of data enhancement,image preprocessing and super-parameter selection,and enhances the image representation ability of images.Compared with the original training method,the learning efficiency is improved,the network convergence speed is accelerated,and the generalization ability of the model is strengthened.Finally,the recognition accuracy rate on the LFW open face test set reaches 96.48%.(3)Using the improved VGG-16 network algorithm model,a real-time face recognition system is designed and implemented,and the performance of the system in the real scene is tested and analyzed by self-built face database.The experimental results show that The system has better recognition effect in the case of changes in expression,posture,illumination,etc.,further verifying the practicability and effectiveness of the improved algorithm. |
