在图像特征提取中, 样本标签并非完全真实有效, 可能导致图像归类框架的分类精度大幅下降, 而现有标签恢复算法面临含噪样本难以高效再利用的瓶颈问题。为此, 本文提出一种基于子集划分迭代投影集成的标签恢复算法。该算法首先随机多次地提取小规模子集信息, 然后综合主成分分析、邻域图正则化及 K-近邻算法等技术进行样本图像的可靠降维与迭代投影集成, 最后遵循多数投票原则实现标签复原。本文选取两大代表性的人脸数据库, 对多种标签恢复算法进行了不同指标下的大量对比分析。实验结果证明, 本文算法能够有效地校正样本的含噪标签, 在同一图像归类框架下针对 Yale B与 AR数据库分别使分类精度提升了 16. 9%与 8. 1%。相较于目前最好的标签恢复算法, 本文子集划分迭代投影集成算法可以提升 4. 3%~4. 7%的分类精度, 且在确保样本数据完整性的同时具备了一定的可扩展性。

In image feature extraction, sample labels are rarely completely true and effective. This often leads to a significant decrease in the accuracy of an image classification framework. In addition, existing label recovery algorithms often must deal with a bottleneck problem in which noisy samples are difficult to reuse. Therefore, this paper proposes a subset-divided iterative projection bagging algorithm for noisy-label recovery. First, the proposed algorithm extracts small-scale subset information randomly and re. peatedly. It then integrates principal component analysis, neighbor graph regularization, K-nearest neigh. bor, and other techniques to achieve effective dimension reduction and iterative projection integration of sample images. Finally, class-label recovery is conducted by implementing the majority voting principle. This study uses common databases as experimental objects and conducts several comparisons and analy. ses of various recovery algorithms using different indicators. Experimental results show that the proposed algorithm effectively corrects the noisy labels of samples, and the classification accuracy of the default framework is improved by as much as 16. 9% and 8. 1% for the Yale B and AR databases, respective. ly. Compared with the state-of-the-art algorithm, the classification accuracy of the proposed algorithm is improved by 4. 3-4. 7%. The proposed algorithm also has good scalability and can ensure the integrity of sample data.