Research On Face Recognition Algorithm Based On Feature Encoding And Unsupervised Learning Mechanism

With the rapid development of Internet of Things and artificial intelligence technology,efficient and convenient identification and authentication technologies have great theoretical research and market application value.Because of its uniqueness,nonreplicability and non-concealment,face recognition have become an important research direction in the field of biometrics.In the actual application,face recognition faces many challenges due to the flow of people and complex scenarios,including: 1)single sample;2)partial occlusion;3)uneven illumination;4)facial expression changes,more robust and more efficient face extraction design is the key to solve these problems.Based on feature learning,this thesis studies the problem of face recognition in complex scenarios,especially single sample,partial occlusion and illumination change.Based on the in-depth study of the theory of feature coding,the innovation and improvement of the existing algorithms are proposed.The main research contents of this thesis are:(1)The multi-scale feature learning method is studied.The characteristics of the traditional feature coding methods are based on a single scale of feature extraction and information expression ability is weak.While multi-scale method has good expression ability for complex image information.So we further research the characteristics of the principle of feature coding and the distribution of the main texture information in face image to propose multi-scale feature coding method.(2)The basic principle of auto coder algorithm and related derivation algorithm are studied.Comprehensive analysis of the advantages and disadvantages of the traditional auto coder algorithm to make improvements.We proposed a combined feature coding model based on unsupervised learning mechanism to improve the efficiency of the feature extraction of the single sample.(3)The feature classification method based on nonlinear mapping is studied.The texture information of face image information in different regions varies widely,and the contribution of each feature to the final recognition is different.Moreover,in complex scenarios face image is often accompanied by partial occlusions,block information loss and error identification on its estimates of interference,this thesis combined with the local structure identifying information to propose nonlinear feature weighting algorithm method to improve the recognition accuracy of complex scenarios.(4)Based on the research of theoretical algorithm,the application of face recognition system is developed.Based on Visual Studio and Android Studio,this thesis realizes the primary development of face recognition system on PC and mobile terminals.