The Research Of Image Recognition Method Based On Tensor

Borned in the1920s, pattern recognition is concerned about the process how to usecomputer algorithm to find the rules hidden in the object or phenomenon automatically, andhow to use these rules to describe, identify, classifiy and interpret the object or phenomenon.As an important research field of pattern recognition, feature extraction not only extracts themost favorable characteristics for pattern classification from the original mode, but alsoreduces the dimension of the original model samples greatly. Pattern recognition can beusually divided into supervised and unsupervised classification. As the crucial step of patternrecognition process, Feature extraction can be divided into feature extraction withdiscriminant information and feature extraction without discriminant information. The methodof feature extraction with discriminant information is often required to provide the class oftraining sample during the training process, however, this is will be difficult in practicalproblems, so depth research on feature extraction without discriminant information becomesvery necessary. In the real world, many objects can be represent as the tensor, for exampleface images can be represent as second-order tensor, silhouette sequence of the humanbehavior can be represent as third-order tensor. And as a result of feature extraction methodbased on tensor can maintain the spatial and temporal characteristics of the original sampleduring the feature extraction process, tensor analysis has become a hot research field incomputer vision and feature extraction. This paper based on the application of behaviorrecognition and face recognition, in-depth study on the tensor-based image feature extractionmethod without discriminant information.This paper first introduces the significance of feature extraction and present situation offace recognition and behavior recognition. Then introduce the related knowledge of tensorrepresentation, the related concepts of which are decrypted vividly and intuitively with thehelp of image. With the rapid development of science and technology, the technology of datacollection are also developed followed therefore, more and more people concerned on how todig out the most useful information from the data, and statistical modeling is undoubtedly oneof the most effective means to deal with this issue. This paper uses the example of diabetes tolead to a linear regression model, and in-depth discussing the advantages and disadvantages offour regression models: least squares method, the edge regression, lasso regression, elastic net.And the respective implementation steps are given.By reading a lot of literature, we found that the dimensions of the silhouette extractedfrom the video action sequences are so high and that it’s very necessary to feature extraction(dimensionality reduction) from the video action sequences. In this paper, we make use of the tensor principal component analysis into the feature extraction of original sample to extractthe most useful information for classification of the original sample. Moreover tensorprincipal component analysis algorithm is a supervised feature extraction method whichneedn’t the category of training samples during the feature extraction process. Then measurethe similarity between the two projection tensor using tensor distance degree. Then constructsa nearest neighbor classifier based on the distance of the tensor to identify the behavior to betested. Experiment results on the famous Weizmann human behavior database shows that ourmethod can not only achieve a higher recognition rate, also can achieve high recognition rateand robustness for different gait.Given a strong noise pollution or reduced face images we can identified them easily,which shows that we needn’t perceive all the image pixels, instead just need to percept part ofthem which can be able to complete the identification task. This means that the human visualsystem has the features of image sparse representation. To this end, this paper proposed asparse tensor feature extraction method based on tensor principal component analysiscombined tensor represented with sparse representation--sparse tensor principal componentanalysis. By transforming eigenvalues problem and eigenvectors problem of the MPCAalgorithm into a series problem of linear regression, this method not only making theprojected tensor as far as possible but also making the projection matrix is sparse. Largenumber of experiments on the AR face database shows that The STPCA algorithm can moreeffectively reduce the effect of occlusion on feature extraction than the original MPCAalgorithm, and DATER and the GTDA algorithm, which both are with discriminantinformation.