Image Feature Extraction Methods Based On Subspace Projection

The feature extraction problem of multi-view data and high-dimensional image data is an important problem in image processing.The image feature extraction methods based on subspace projection have advantages of low computational cost and strong description ability.And these methods are widely used in multi-view image recognition and dimensionality reduction.Therefore,the paper topic is devoted to investigate the problem that the existing multi-view common subspace projection methods can not make full use of the effective information in multi-view data during the subspace learning process,and the problem that the existing low-dimensional subspace learning methods pre-learn the reconstruction relationship.In this paper,the existing algorithms are improved to enhance the feature extraction performance for multi-view image data as well as high-dimensional image data.The main work of the thesis includes:(1)Based on the multi-view common subspace projection methods,this paper proposes a weighted multi-view common subspace projection method,which effectively adjusts the contribution of intra-view and inter-view discriminative information by the weighted parameter in order to remove noise and redundant features and make full use of the multiview information to extract effective features.To solve this model,the maximum scatter difference criterion is utilized as a metric for intra-view and inter-view projections.The proposed method is experimented on three datasets,Caltech101-7,Caltech101-20 and MSRCV1,and eight existing related multi-view common subspace projection methods are also compared on the same datasets.The experimental results show that the average recognition accuracy of the proposed weighted multi-view common subspace projection method can be improved by about 5%.(2)Based on the low-dimensional subspace projection methods,this paper proposes an adaptive neighborhood preserving discriminative subspace projection method,which automatically updates the sparse reconstruction coefficient by the designed algorithm in the process of subspace learning and also learns the fuzzy membership relationship between data points and clustering prototypes in the low-dimensional subspace.Through the integration of adaptive reconstruction relationship and fuzzy projection clustering,the proposed method can explore both reconstruction relationship and clustering structures.In addition,an iterative algorithm to alternatively update the variables in the method is designed in this paper for solving the method.The proposed method is experimented on four datasets,MSRA,Dermatology,Satimage,and Yeast.And seven existing low-dimensional subspace projection learning methods are also compared on the same datasets.The experimental results show that the proposed adaptive neighborhood preserving discriminative subspace projection method can improve ACC values by 1%～13% and NMI values by 1%～17%.(3)The proposed weighted multi-view common subspace projection method and adaptive neighborhood preserving discriminative subspace projection method are applied to face recognition,and experiments are conducted on the CMU-PIE multi-view face dataset and three single-view face datasets,including the ORL dataset,LFW dataset and Yale dataset.The experimental results show that both the proposed weighted multi-view common projection subspace method and the adaptive neighborhood preserving discriminative subspace projection method can effectively extract features in face images on the above face datasets.The proposed weighted multi-view common subspace projection method can make full use of multi-view information by adjusting the contribution rate of intra-view and inter-view discriminative information,and effectively extract the features of multi-view data.The proposed adaptive neighborhood preserving discriminative subspace projection method can learn the reconstruction relationship and clustering structure in high-dimensional data,and effectively extract features in high-dimensional image data.Finally,the proposed two subspace projection image feature extraction methods are applied to multi-view face recognition dataset and single-view face recognition datasets,the experimental results demonstrate the effectiveness of both methods.