Sample Label-based PLS and Feature extraction

摘要

Partial least squares (PLS) method is an effective approach for regression analysis and image feature extraction. Non-iterative PLS based on orthogonal constraints can extract PLS features rapidly and effectively, while the features maybe correlative. PLS based on Uncorrelated Score Constraints can extract Uncorrelated features which make image recognition more effective. 2DPLS can extract features from image matrices directly, which can solve the small sample problems. Considering that the traditional class label encodings don't emphasize the significance of the samples in overlapping regions between classes, fuzzy k-NN method is employed in class label encodings to make use of the sample distributions, then improved algorithms of PLS and 2DPLS based on sample label encodings are given. The results of experiments on ORL face database show that the improved algorithms presented are better than traditional PLS, which can extract discriminative features more efficiently and robustly.