Abstract: Focusing on the issues that the sparseness of data is reduced after factorization and the single image feature cannot describe the image content well, a multi-constraint nonnegative matrix factorization based on feature fusion was proposed. The information provided by few known labeled samples and sparseness constraint were considered, and the graph regularization was processed, then the decomposed image features with different sparseness were fused, which improved the clustering performance and effectiveness. Extensive experiments were conducted on both Yale-32 and COIL20 datasets, and the comparisons with four state-of-the-art algorithms demonstrate that the proposed method has superiority in both clustering accuracy and sparseness.

Key words: Non-negative Matrix Factorization (NMF), label information, sparseness constraint, graph regularization, feature fusion

摘要： 针对非负矩阵分解后数据的稀疏性降低、单一图像特征不能够很好地描述图像内容的问题，提出一种基于特征融合的多约束非负矩阵分解算法。该算法不仅考虑了少量已知样本的标签信息和稀疏约束，还对其进行了图正则化处理，而且将分解后的具有不同稀疏度的图像特征进行了融合，从而增强了算法的聚类性能和有效性。在Yale-32和COIL20数据集上进行的对比实验进一步验证了该算法具有更好的聚类精度和稀疏性。

关键词: 非负矩阵分解, 标签信息, 稀疏约束, 图正则, 特征融合