PROJECT TITLE :

Low Rank Embedding For Robust Image Feature Extraction - 2017

ABSTRACT:

Robustness to noises, outliers, and corruptions is a crucial issue in linear dimensionality reduction. Since the sample-specific corruptions and outliers exist, the class-special structure or the local geometric structure is destroyed, and thus, many existing methods, together with the popular manifold learning- based mostly linear dimensionality strategies, fail to realize sensible performance in recognition tasks. During this paper, we tend to concentrate on the unsupervised sturdy linear dimensionality reduction on corrupted information by introducing the strong low-rank representation (LRR). Therefore, a strong linear dimensionality reduction technique termed low-rank embedding (LRE) is proposed during this paper, that provides a robust image illustration to uncover the potential relationship among the pictures to scale back the negative influence from the occlusion and corruption therefore as to reinforce the algorithm's robustness in image feature extraction. LRE searches the optimal LRR and optimal subspace simultaneously. The model of LRE can be solved by alternatively iterating the argument Lagrangian multiplier methodology and the eigendecomposition. The theoretical analysis, including convergence analysis and computational complexity, of the algorithms is presented. Experiments on some well-known databases with completely different corruptions show that LRE is superior to the previous strategies of feature extraction, and so, it indicates the robustness of the proposed technique. The code of this paper can be downloaded from http://www.scholat.com/laizhihui.