Research On Image Recognition Method Based On Supervised Multi-manifold Learning

In the era of big data,it will help to deal with and apply big data through effectively collecting and analyzing massive data to discover the essential features hidden in the high dimensional data.Manifold learning can map the high dimensional data into the low dimensional space and find the essential characteristics of the high dimensional data,which overcomes the curse of dimensionality.But most manifold learning methods don't provide an explicit function between two spaces,so that they are unable to directly reduce the dimension of data out of the training set.And it is not suitable to classifying samples with the low features because the main purpose of manifold learning is to visualize the high dimensional data.Some methods,such as LPP,MMDA and so on,assume that the high dimensional data are located in multiple sub-manifolds.And samples from different classes are distributed on various sub-manifolds.These methods are applied to image classification and have achieved better results.However,some of them ignore the class labels or some of them appear singular matrix,which greatly affects the recognition effectiveness.To solve these problems,a classification method based on supervised multiple manifold learning is proposed,which is applied to image recognition and achieves better results.The main work of this paper is as follows:(1)A classification method based on supervised multi-manifold learning was presented,which utilized the class labels.In the algorithm,the projection matrix was obtained through maximizing the Laplacian diagram between classes.The method is a supervised linear method.The method was tested and compared with other image recognition methods on the pavement damaged image database,ORL and FERET.Experimental results showed that the proposed method greatly improved the accuracy of image recognition and overcame the problem of the singular matrix in MMDA and LPP.(2)The method of choosing the dimension of the low dimensional space was studied.According to the idea of PCA to determine the low dimensional dimension,a proper low dimensional dimension is selected through finding out the relationship between the cumulative contribution rate of the eigenvalues and the recognition accuracy.Experimental results on face databases demonstrated that a proper low dimensionaldimension could be determined through the proposed method.(3)The effect of different similarity measures on the accuracy was analyzed and compared.In the experiment,two similarity measures,Euclidean distance and the angle cosine distance,were compared.Experimental results showed that various data sets need to adopt different similarity measures.Even if for the face databases,better results were gained from Euclidean distance on ORL face database.However,better results were obtained from the angle cosine distance on FERET face database.