Research On Unconstrained Face Recognition Based On DBNs Network

The unconstrained face recognition has been a core content in the field of personal identification because of high concealment,non invasive and practicality.The unconstrained face images have characteristics of high degree of freedom and disturbance of complex feators,and normal algorithms can not make a perfect identification effect.In recent years,some new artificial intelligent aigorithms have been successfully used in the field of classification and targets identify,and that is represented by the theory of deep learning.For this reason,the paper designs the algorithm of unconstrained face recognition based on DBNs model and researchs the generality in other areas.Firstly,the paper describes and analyzes the difficulty about unconstrained face recognition,and summarizes the usual algorithms used currently.Secondly,the paper researchs the theory of deep learning and designs the algorithm of unconstrained face recognition based on DBNs model successfully.To improve the performance of algorithms,the paper chooses Relu as activation function and designs the optimization algorithms based on KL relative entropy sparse restrictions and dropout mechanism.Then,the paper proposes a algorithm based on hybrid DBNs model,which can generate simulated samples to train the DBNs model with CNNs model.In order to ensure the new samples have a good diversity and difference,the new model changes convolution kernel and the mapping relationship between sub-sampling layer and convolution layer in training step.Finally,the paper proves the generality of algorithm model in cifar-10 and SAR data set.For feature enhancement,the paper uses two-dimensional entropic thresholding to partition SAR images into shadow and target in the pre-processing stage.The experimental results show that the algorithm can fully satify the requirement of practical applications and keep a good recognition rate in the situations of lack sample sizes.The algorithm model also has good generality and can serve as references for other researchers to solve similar problems.