Research On Face Recognition And Dimension Reduction Of Stacked Autoencoder

With the vigorous development of internet technology,the application of machine learning and deep learning in image,language,vision and other fields has been particularly prominent.However,high-dimensional data restricts machine learning,which greatly increase the time and space complexity of the model.In face recognition,the original high-dimensional images contain a large amount of noise and redundant information,which reduce the recognition rate of the classification model.Therefore,it is especially important to map high-dimensional data into low-dimensional space and fully express the internal structure of data.Traditional linear data dimensionality reduction methods are limited to low-dimensional data,non-linear dimensionality reduction algorithms have problems such as non-convergence or sensitivity to noise.The stacked autoencoder(SAE)is a branch structure of deep learning.It can self-learn image features by layer-by-layer training and capture the most significant features in training samples information.At the same time,SAE is less complete autoencoder,this special structure can enable high-dimensional data to achieve dimensionality reduction without losing important information.Therefore,this paper proposes an algorithm to realize the face recognition for dimension reduction by stacked autoencoder structure.Firstly,the face images are preprocessed,and the profile information is roughly extracted through the second generation discrete curvelet transform and then input as curvelet faces into the stacked autoencoder,and the network is trained with feature vectors of the curvelet faces,the labeled data samples are used to fine-tune the network to achieve the optimal network.Finally,the multi-layer features of SAE are combined to obtain richer classifier training data.Experimental results show that the stacked autoencoder has good learning characteristics and dimensionality reduction ability,also has stronger robustness against small sample data,the recognition accuracy in low-dimensional space is improved compared to the traditional methods.At the same time,the curvelet features can reduce the complexity of the SAE,including rich edge details while removing irrelevant redundant features.Finally,the SAE feature fusion algorithm has higher recognition accuracy,which verifies the effectiveness of the algorithm.