Abstract: Principal component analysis algorithm (PCA) is one of the classic algorithms for feature extraction and denoising. Its improvement has always been a hot topic in the field of feature extraction learning. As an improved algorithm of robust principal component analysis, locally invariant robust principal component analysis (LIRPCA) has better learning performance and improves the efficiency of feature extraction compared with most existing matrix-based subspace learning methods. However, the category relationship between samples is not considered in this algorithm, so it belongs to the unsupervised learning algorithm, and its performance in recognition performance is relatively poor compared to the supervised learning algorithm. In order to improve the recognition performance of the algorithm as much as possible, a feature extraction technology based on nearest neighbor supervised local invariant robust principal component analysis (NSLIRPCA) is proposed here.. The NSLIRPCA algorithm is based on the existing LIRPCA feature extraction algorithm model and has made corresponding improvements. Based on the former algorithm, the category information between samples was added, and the relationship matrix was constructed based on this. Then, the algorithm will solve the formula and prove the convergence of the formula. Finally, it is found that the recognition rate of the algorithm is improved compared with other classical algorithms by applying the algorithm to various occlusion databases.

Key words: Keywords: feature extraction, subspace learning, unsupervised learning, robustness, image identification

摘要： 主成分分析算法(PCA)作为特征提取去噪的经典算法之一，对于它的改进一直以来都是特征提取学习研究领域的关注热点内容。局部不变鲁棒主成分分析（LIRPCA）作为鲁棒PCA的改进算法，它与大多数现有的基于矩阵的子空间学习方法相比，具有更好的学习性能，提高了特征提取的效率。但是，该算法没有考虑到样本间的类别关系，属于无监督学习的算法，相对于监督学习的算法来讲，识别性能较差。为了尽可能地提高该算法的识别性能，在此提出了一种基于近邻监督局部不变鲁棒主成分分析（NSLIRPCA）的特征提取技术，NSLIRPCA算法是在现有的LIRPCA特征提取算法模型的基础上做出了相应的改进变换，基于前者的算法基础上又增添考虑了样本间的类别信息，并以此来构建关系矩阵，接着再对该算法进行公式求解和公式的收敛性证明，最后通过将该算法应用于各种遮挡数据库中发现，相对于其他经典算法，该算法识别率均有所提高。

关键词: 关键词: 特征提取, 子空间学习, 无监督学习, 鲁棒性, 图像识别