摘要： 针对无标签样本和单标签样本的融合学习问题,提出样本稀疏邻域的概念,进而给出基于稀疏邻域的特征融合算法(SNSPDA)。样本的稀疏邻域充分利用稀疏表示的判别属性,增强了具有较大表示系数样本对被表示样本的重构作用。SNSPDA算法可捕获数据的局部几何结构,保持样本间的稀疏重构关系,同时避免单标签样本学习中的过拟合问题。大量单标签图像样本的实验结果表明,SNSPDA算法比仅反映单一数据属性的融合算法具有更高的识别率,如在光照条件变化较大时,该算法的正确识别率分别比稀疏保持判别融合算法与半监督判别融合算法提高了2.14%与17.43%。

关键词: 特征融合, 稀疏邻域, 正则化, 几何结构, 稀疏重构, 特征分解

Abstract: Concerning the fusion learning problem of the unlabeled and single labeled samples,this paper gives the concept of samples sparse neighborhood,then further puts forward Sparisity Preserving Discriminant Analysis Based on Sparse Neighborhood(SNSPDA) algorithm.Samples sparse neighborhood makes full use of its discriminant attribute,and SNSPDA reinforces the role of those samples which have big reconstructive coefficients.This algorithm not only captures the local geometry structure,but also maintains the sparse reconstruction relationship between samples.Furthermore,it avoids the overfitting problem during the process of the single labeled sample learning.A mass of experimental evidence from single labeled image samples demonstrates that this fusion feature algorithm has a higher discriminating rate than those fusion methods which only reflect single data attribute.For instance,when the illumination condition changes significantly,the distinguishing rate of SNSPDA raises by 2.14% and 17.43% compared with Sparsity Preserving Discriminant Analysis(SPDA) algorithm and Semi-supervised Discriminant Analysis(SDA) algorithm.

Key words: feature fusion, sparse neighborhood, regularization, geometric structure, sparse reconstruction, eigen-decomposition

中图分类号: 

• TP393