图像分类中多流形上的词典学习

传统的基于稀疏编码的图像分类算法没有考虑不同特征之间的关系。针对这一不足,该文提出了一种新的词典学习算法。该算法考虑特征所处的多个流形空间上的拓扑结构,并显式地对其进行建模,然后将该模型嵌入到稀疏编码算法中构造多流形上的词典优化目标函数。为求解该优化问题,还提出了使用坐标下降的方法,同时给出了收敛性分析。在图像分类3个基准数据集上的实验结果表明,提出的算法分类性能超过了基于传统稀疏编码的算法。这也进一步证明了该算法的有效性。

Dictionary learning on multiple manifolds for image classification

Traditional image classification algorithms based on sparse coding do not consider the relationships among different features.A dictionary learning algorithm was developed to overcome this shortcoming,in which the intrinsic geometrical structure of the multiple manifolds explicitly models the features embedded into a sparse coding algorithm.A coordinate descent algorithm was then used to solve the optimization problem.A convergence analysis was given that shows the algorithm will converge.Tests on three standard benchmark datasets show that the algorithm outperforms other state-of-the-art image classification algorithms.