Research On High-order Multi-attribute Signal Recognition Method Based On Wavelet Transform And Tensor Decomposition

With the progress and development of information technology these days,the unstoppable development of information society makes the demand of information process higher and higher,the traditional information process method can't meet the needs,so how to use machines to carry on information process and big data analysis is becomes a new trend.How to process massive and complex information effectively and quickly and extract the needed information as much as possible has become a hot topic direction nowadays.At the present,most of the method of information process is only processing the single attribute of data,and most of the method process method of multi-attribute having on hand is extracting the feathers from different feathers of attribute and then join on or merge them,this method has the lack of destroying the structural characteristics of multi-attribute data feathers for the different data which has correlation between different attribute.Besides,for the nowadays feathers extracting and recognizing method process are almost carry on at time domain,and neglect some feathers at frequency domain,so this article study this problem,and chose the wavelet domain as the main study direction after the comprehensive comparison,and combine the neural network and wavelet transformation for research and study it.Convolution neural network has achieved good effect at image classification and voice recognition,but with continue claim of data mass and data analysis capability,single attribute analysis to image and other low-dimensional analysis can no longer meet the needs.CNN is good at capturing the spatial characteristics of data,and frequency domain analysis is good at capturing scale invariance based on spectrum information.This article combines the two with tensors,and express the data with multi-attribute characteristics with tensors.The data is expanded into the wavelet domain and combined with high-order convolutional neural networks to classify multi-attribute data.Compared with the traditional CNN,it can improve the efficiency and lift the classification accuracy.This article's research can be summarized as follows:1.Propose a classification algorithm merging wavelet transformation with convolution neural network image.The convolution neural network processes and analyzes the data characteristics through multi-layer and non-linear information processing,from some angles,CNN can be seen as one of multi-resolution analysis,and wavelet transformation has the characteristics as well.However,CNN is hard to approximately learn the filter invariant parameters in wavelet transformation from the data,so this article proposes an algorithm after extracting the wavelet domain characteristics,and input them into CNN and carry on learning classification algorithm.This method has conducted experiments on multiple data sets and achieved good results..2.Propose a classification algorithm merging wavelet transformation and high-order convolution neural network with multi-attribute signal.Through tensors to express model expand the convolution neural network to the tensor space,and use wavelet transformation to analyze the time-frequency domain characteristics of data and obtain higher precision and effective information processing.At the same time expand the wavelet domain information to tensor space to get more comprehensive learning multi-attribute characteristics of data,and define tensor convolution operation,tensors pooling operation,high-order forward propagation and high-order back propagation algorithm to train the parameters in convolution neural network in high-order tensor space,3.Implemented multiple application examples of high-order multi-attribute signal classification of tensor wavelet.Experiments were carried on multi-attribute datasets such as Poly-U,CAUVE,MIR Flickr and other multi-attribute data sets to verify the performance of the proposed high-order convolution neural network algorithm merging with tensors wavelet.The experiments result shows that the high-order neural network algorithm of merging tensors wavelet in this article has higher classification accuracy than the depth calculation model and multi-mode model in big data,and it verify the feasibility of the proposed merging method proposed in this article.