Research And Implementation Of Face Recognition Based On Triplet-awared Center Loss

Face detection refers to a series of computer technology that is able to identify people's faces within digital images.Based on face detection,face recognition is is able to give the corresponding identity information of faces.With the rapid development of deep learning,face recognition has achieved a great leap-forward improvement in recognition accuracy and application reliability.And it has been widely applied to company attendance,station automatic ticket checking,APP payment authentication and other scenarios.In the field of computer vision and pattern recognition,the performance improvement of the deep learning methods benefits from the massive data,enormous computing power of GPU,the deep network structure and the various network optimization methods.As a branch of the field,face recognition also needs these necessary conditions to improve its accuracy.In this thesis,a recognition architecture based on the Residual Network is proposed to train a model on MS-CELEB₁M,the millions of celebrity data set published by Microsoft.Then,the proposed recognition architecture is tested the model on some international standard testing data sets,such as LFW,YTF,CFP,etc.The main work of this thesis is as follows:1.For MS_CELEB₁M training set contains a lot of noising images and can not be labeled manually,a cleaning method based on visual similarity is proposed in this thesis.First an existing deep model is used to extract the deep features of each category.Second the K-means clustering algorithm is used to divide all the features into two groups,each of which has a centroid,and select a centroid of one of the group which has more images as the real centroid.Then the cosine distance between each image feature and the selected center is calculated,and we keep the image whose cosine distance is in a certain range,otherwise,we delete it.Experiments show that the proposed method can delete effectively noisy pictures and improve training efficiency.2.For extracting discriminative features making the metric distance between the same face small and the metric distance between different faces large,a new loss function called Triplet-awared Center Loss is proposed in this thesis.The method combines the advantages of both the Center Loss and the Triplet Loss,which decreases the distance between same classes while increasing the distance between different classes.In addition,it avoids the disadvantage of Triplet Loss that requires careful selection of Triplet-sets,which makes the training phase more concise and effective.3.From the perspective of deep convolution neural network,the influence of the size oftraining set,the depth of the network and the SE network based on the correlation of the channel dimension on the face recognition is explored.Extensive experiments show that the Triplet-awared Center Loss proposed in this thesis can converge stably to the best state when the training set is large,and the network is deepened or changed.It shows that the loss function proposed has a strong robustness under the given optimal parameters??=0.1,?=20,K=20?.Besides,the influence of occlusion on face recognition is explored and it is found that occlusion at nose part has the greatest impact on accuracy.The accuracies of LFW,YTF,CALFW and CPF-FP test set are increased to 99.62%,95.94%,94.68%,94.99% respectively.Especially for BLUFR,the face recognition accuracy at DIR@FAR=1% is improved from 69.79% to 92.57%.