带判别性局部结构约束的多分辨率字典学习算法及人脸识别研究

Multi-Resolution Dictionary Learning Algorithm with Discriminative Locality Constraints for Face Recognition

字典学习是图像表示的一种有效手段且在图像识别任务中具有出色的表现。大部分传统字典学习算法在多分辨率人脸图像识别中存在不同分辨率下字典判别性不强的问题。为了解决这一问题,提出一种新的带判别性局部结构约束的多分辨率字典学习算法。首先,通过分析字典原子与轮廓向量之间的一一对应关系,采用字典原子的局部几何结构,分别构造了轮廓向量的类内局部约束项与类间局部约束项;然后将这两个约束项统一在同一个正则项中,并将其增加到字典学习目标函数中进行联合优化,从而实现判别性局部几何结构的编码。该算法促使类内编码系数保持相似的局部一致性,而且能有效提高类间编码系数的局部结构的差异性。最后,在多个多分辨率人脸图像数据集上验证了本文算法的有效性,实验结果表明,与同类字典学习算法相比,本文学习的多分辨率字典能保持训练样本中的判别性局部结构,在不同分辨率的人脸图像上获得了更好的识别性能。

Although dictionary learning has shown to be a powerful tool for image representation and has achieved satisfactory results in various image recognition tasks.Most traditional dictionary learning algorithms have been restricted to multi-resolution face recognition tasks mainly due to the poor discriminability of the dictionary.To solve this problem,we propose a novel multi-resolution dictionary learning algorithm with discriminative locality constraints(MDLDLC)in this paper.Based on the one-to-one mapping between each dictionary atom and the corresponding profile vector,we design two local constraints on profile vectors,referred to as intra-class and interclass local constraints,by utilizing the local geometric structure of the dictionary atoms.Meanwhile,the two constraints are formulated into a unified regularization term and incorporated into the objective function of the dictionary learning model to optimize for encoding the discriminative locality of input data jointly.The proposed MDLDLC algorithm encourages high intra-class local consistency and inter-class local separation in the code space of multi-resolution images.Finally,extensive experiments conducted on different multi-resolution face image datasets demonstrate the effectiveness of the proposed MDLDLC algorithm.The results show that the proposed MDLDLC algorithm can learn the multi-resolution dictionaries with discriminative locality,preserving and achieving promising recognition performance compared with other state-of-the-art dictionary learning algorithms.