Research And Application Of Face Recognition Method Based On Unconstrained Conditions

Currently,deep convolutional neural networks are widely used in the field of computer vision as their distinguished feature extraction function,and the method of using deep convolutional neural networks is the current mainstream study orientation,especially in the field of face recognition.In this paper,we conducted a study of face recognition algorithms under unrestrain environment on account of deep learning,and improved the softmax loss function and cosine interval loss function.The specific research contents are as follows:(1)A recognition algorithm for unconstrained faces is proposed,and the primitive loss function is improved to lower the impact of disequilibrium training caused by lack of data example under unrestrain environment.The examples with lack of data are referred to the idea of triples.Realize the expansion of the data,and secondly,by introducing additional parameters,decide the degree of distinction between the data sample and the negative sample,so decide whether to increase the degree of discrimination for the sample loss.The experimental part uses MTCNN as the detection network to geometrically correct the key points of the detected face.The processed data images can be used for training and testing.Finally,the improvement is verified on the cross-age test set Age-DB and the side face image test set CFP After the model.Experimental results show that the improved Softmax has good practicability for face recognition under unconstrained conditions.(2)Aiming at the million-level face recognition needs,a recognition algorithm based on the cosine angle interval loss function is proposed.Binary parameters are introduced under the original loss function to emphasize the distinction feature learning of the misclassified trait vector,and the training process is realized.Adaptive change.The experimental part uses the million-level test set LFW and MegaFace to test the improved algorithm,and the characteristic extraction model uses the Inception-ResNet construction network.The experimental results show that the improved angular interval loss function is for people under the million-level sample.Face recognition not only improves the recognition accuracy,but also has good robustness for unconstrained face recognition.(3)A face recognition institution based on convolutional neural network is established.Use PyQt5 to design GUI reciprocal interface,which is convenient for the input of face message and the demonstrate of face recognition effect.And the recognition system was measured on the face occlusion and face attitude on the self-built small database to verify the expression of the system.