Research Of Metric Learning Algorithm Based On Local Structure Preserving

In the fields of computer vision and pattern recognition,similarity metric learning is the most basic method to measure the similarity among the sample feature vectors.Learning a good similarity metric function effectively by using prior information has become one of the research hotspots.At present,many metric learning algorithms,which learn the metric matrix by minimizing the intra-class distance and maximizing the inter-class distance,have been proposed.And the metric learning algorithms based on Mahalanobis distance are the most commonly used.However,there are still some problems such as the lack of the incorporation of global and local information,the lack of robustness in a noisy environment.To solve these problems,we propose three different metric learning models based on local structure preservation.The main achievements are summarized as follows:(1)In order to make full use of the label information of training samples,a robust metric learning algorithm based on label regression is proposed.The algorithm obtains a discriminative latent metric subspace by fitting the binary label matrix.Meanwhile,the local data structure is mined by introducing the sparse representation learning.Then,the "clean" representation of the original data in the subspace is obtained.In addition,the nuclear norm is used to measure the regression loss to improve the robustness under noise conditions.The experimental results on several datasets show the effectiveness of the algorithm.(2)In order to solve the problem of information loss and overfitting in LDA-based algorithms,a regularized metric learning algorithm based on bidirectional reconstruction constraint is proposed.Firstly,we propose a bidirectional reconstruction constraint which uses an orthogonal matrix different from the projection matrix to reconstruct the data,so that the low dimensional feature transformation matrix is given higher degrees of freedom to enhance the flexibility of subspace dimension selection.Secondly,the proposed method uses the class compactness graph to preserve the local data structure,thereby avoiding the problem of overfitting.The effectiveness of the proposed similarity metric algorithm is proved by the classification results on several datasets.(3)In order to address the problem that the noise in the original data imposes a negative influence on the construction of similarity matrix,a self-weighted local preserving discriminative metric learning method is proposed.Firstly,this model adjusts the local relationships of within class samples by calculating the weighted matrix automatically.And the algorithm calculates the corresponding self-weighted matrix in each optimized metric space,which can weak the influence of noise in the original data space.Secondly,low-rank basis joint learning is introduced to capture the global data structure and improve the anti-noise ability of the algorithm.Finally,a Mahalanobis distance metric function incorporating local and global structural information is obtained.The effectiveness of the proposed algorithm is verified by comparing with the mainstream algorithms on several datasets.