Research And Implementation Of Face Recognition Algorithms For Video Sequences Under Nonideal Conditions

As a research direction of computer vision,biometric technology has been highly concerned by domestic and foreign researchers,enterprises and institutions.Face-recognition technology is developing rapidly because of its high commercial value and extremely broad application prospects,which highlights its importance in the direction of biometric technology.However,although face-recognition technology has been developed for decades,there are still many technical bottlenecks that have not yet broken through due to the application of complex real-world scenarios.In this paper,according to the systematic operation process of face recognition,some key technologies of face recognition are surveyed.The CASIA-WebFace data set is selected as the training set,some key points of face are collected to judge the attitude angle of face by the transformation of quaternion and Euler angle.The face image satisfying the constraints is screened,then the training set is expanded with small samples by the image multi-scale transformation and normalized.By combining the advantages in the deep convolutional neural network and the residual network,the network is optimized as zxzres-cnn network.AM-Softmax is used as the objective function to train the pre-processed training set.The problems of over fitting,gradient vanishing and depth model degradation are avoided.The validation set is used to test the model,and the highest recognition rate is 99.9709%.After training,the face video sequence images under nonideal conditions are collected and inputed into the network for testing.The face images are compared using the cosine similarity of angles.Finally,the zxzres-cnn algorithm is successfully applied to the visualization program in Python and C programming language.Then the face recognition system is tested.Experiments show that the system works well for face recognition under nonideal conditions.