Face Recognition Method Based On Attention Mechanism And Curriculum Learning

As an important biometric recognition technology,face recognition is widely used in finance,transportation,military and other important field.Deep face recognition obtains more robust face representation by training convolutional neural network with large-scale data sets,greatly improving the performance of face recognition.At present,deep face recognition mostly relies on convolution neural network to extract high semantic face features,so designing effective convolution network model has become an important research direction.The rapid development of convolution neural network model further improves the performance of deep face recognition,making the accuracy of face recognition nearly saturated.In recent years,the research on deep face recognition has gradually shifted to the loss function,with the aim of making the face distribution of different identities as scattered as possible,and making the face distribution of the same identity as compact as possible.First of all,analyzing and studying the mainstream attention mechanism,It was found that these attention mechanisms only highlight features of a certain dimension of the image or the module design is relatively complex,in order to compensate for the shortcomings of previous attention mechanisms,further highlight the channel and spatial features of facial images and enable the network model to extract higher quality facial features,an efficient spatial channel attention module(ESCA)is proposed and integrated into the basic module of the feature extraction network.This module uses efficient channel attention(ECA)to obtain channel attention and adds spatial attention module after ECA to further obtain spatial attention on the basis of focusing on image channel information.The network model with different attention modules is trained on the CASIA-Web Face dataset.The experimental results show that adding ESCA module can achieve better recognition performance than adding other attention modules.Integrating the ESCA module into the lightweight network model Mobile Face Net and the shallow network Res Net50,and results show accuracy improvement of over 1.5% compared to the original network.By analyzing the convergence curve of the Curricular Face network model during training and the network model after adding ESCA,it is found that adding ESCA module can effectively accelerate the convergence speed of the network.Secondly,in order to reduce the interference of unrecognized facial images on the network model during the course based learning process and adapt the loss function Curricular Face based on curriculum learning to low-quality face recognition,the Curricular Face+ is further improved,and the Curricular Face+ which integrates feature norm as image quality indicator into curriculum loss function is further proposed.At the same time,in order to make more effective use of training samples and give full play to the advantages of the ESCA module,a face recognition algorithm(Efficient Cooperative Attention and Curriculum Face)called ECACFace,which combines attention mechanism and curriculum learning is proposed.The algorithm uses the loss function called Curricular Face to train the network model integrated with the ESCA module.Three different loss functions which includes Curricular Face,Ada Face and Curricular Face+ are used to train the network model in the experiment.The experiment uses CASIA Web Face as training set to train the network model and test it on IJB-B,IJB-C and Tiny Face.Compared with the model trained by Curricular Face,the verification accuracy of Curricular Face+ on IJB-C is improved by 12.81%,and the recognition accuracy of Rank-1 and Rank-5 on low-quality data set called Tiny Face is improved by 2.66% and 2.2% respectively.In order to further improve the recognition performance of ECACFace for low-quality face images,we use Curricular Face+ to replace Curricular Face in ECACFace,and further propose ECACFace+.ECACFace and ECACFace+ were trained on the millions of MS1MV2,the experimental results show that the fusion of ESCA module and loss function based on curriculum learning can further improve the performance of face recognition.The experimental results on the high-quality face test set show that ECACFace+ maintains the same performance as ECACFace.Compared with ECACFace,the accuracy of Rank-1 and Rank-5 of ECACFace+ on the low-quality data set called Tiny Face has increased respectively.Experiments show that the improved curricular loss function called Curricular Face+ is more effective in improving the performance of low-quality face recognition.