The Research Of Lightweight Face Recognition Algorithm Based On Deep Learning

Thanks to unique physical properties and non-contact characteristics,face recognition technology has become one of the most important identity recognition technologies,and has been widely used in many fields.In face recognition,how to extract high-quality face features is the key to determining the performance of the algorithm.Before the rise of deep learning technology,face recognition algorithms mainly relied on handcraft features carefully designed by experts for feature extraction.Handcraft features are generally designed for certain specific quests or requirements,relying on a priori knowledge in these fields,with high development costs and limited application scenarios.Deep learning technology is a hot topic in the machine learning community.It aims to learn potential patterns and representations directly from samples through complex non-linear structures with multiple cascades,breaking the limitations of handcraft features and implementing automatic feature extraction.In recent years,as one of the representatives of deep learning technology,convolution neural network(CNN)has become a research hotspot in the face recognition community.As the depth and performance of CNN models continue to increase,the computational cost in the training process is also increasing,so that the training of most high-performance models depends on expensive hardware.Therefore,how to realize the lightweight and miniaturization of the model while maintaining the performance of the model is the key to the further development and application of face recognition algorithms based on deep learning.This dissertation focuses on face recognition algorithms based on deep learning as the main research topic,carefully studies their algorithm ideas and shortcomings,analyzes their defects and proposes improved algorithms,so as to improve the performance and robustness of the algorithm.Specific research works are summarized as follows:(1)Principal component analysis network(PCANet)is a lightweight deep learning model,which uses PCA instead of back propagation algorithm to train the model,greatly reducing the model training time.In addition,a typical PCANet model contains only two convolution layers,a binary hash layer,and a histogram statistical layer,which reduces the amount of calculation from the network structure.However,PCANet uses a L2-norm principal component analysis algorithm,which is not robust to outliers and noise.L1-PCA is a simple and efficient L1-norm-based principal component analysis algorithm,which is more robust to outliers.In this paper,L1-norm-based principal component Analysis network(L1-PCANet)is proposed by introducing L1-PCA algorithm into PCANet.In addition,we also introduced a 2D strategy into L1-PCA,derived L1-2DPCA,and introduced it into PCANet,constructing a L1-norm-based two-directional two-dimensional principal component analysis network(L1-2D~2PCANet).We conducted a lot of experiments on multiple face datasets.The experimental results show that the proposed algorithm can effectively improve the performance of the model,and it is more robust to samples with outliers and noise.(2)At present,we can only input a certain test set into the network and observe the output if you want to evaluate the performance of CNNs.In this way,the effect of the evaluation depends on the selection of the test set.When the number of sample sets is too large,performance evaluation will take more time,especially during model training,the performance of the model needs to be evaluated every certain training batch.Singular vector canonical correlation analysis(SVCCA)is a method used to measure the similarity between any two groups of neurons.In this paper,we propose a new method for evaluating the performance of CNNs based on SVCCA.This method measures the known CNN and unknown CNN by SVCCA.We conducted a lot of experiments on multiple classic data sets.The experimental results show that the proposed method can improve the evaluation speed by 5 times when it has an evaluation error of about 5%compared to the traditional evaluation method.PCANet processes all the samples into a sample matrix during training,and perform the PCA operation once to finish the training.When new samples coming,PCANet can only be retrained.In addition,the sample matrix will exceed the hardware limit when there are too many training samples,making the training impossible.PCANet cannot train under a large data set.The cascading covariance-free incremental principal component analysis algorithm is an efficient incremental principal component analysis algorithm.We expand the CCIPCA algorithm to a two-directional two-dimensional incremental principal component analysis algorithm(2D~2IPCA),and then introduces it into PCANet,constructing a two-directional two-dimensional incremental principal component analysis network(IPCANet).The proposed network uses only one sample at a time during the training phase,thus avoiding the situation where the sample set is too large to train.We conducted a lot of experiments on multiple classic face data sets.The experimental results show that the method proposed in this dissertation not only improves performance,but also solves the problem of too large sample sets.