基于鲁棒判别式约束的深度字典学习及人脸识别应用研究

In order to improve the robustness of face recognition, the project focus on how to design the models of deep dictionary learning for enhancing the ability of extracting face features by using the theory of deep learning and dictionary learning. The project will involve: (1) constructing deep dictionary learning model based on the network of auto-encoder by using the analysis dictionary and synthesis dictionary. The model builds up the hierarchical model which extracts the face features from the bottom level to the top level by simulating the hierarchical structure of the human brain. Thus, it can overcome the shorting of traditional dictionary learning method which cannot extract the deep features; (2) designing robust discrimination model by using the line vectors of coding coefficients(Profiles), analysis coefficients and atoms of each level. The model not only can inherit the upper features, but also can preserve the features of this layer. Thus, the relations among the nodes, weights and deep features can be established, and it can improve the performance of feature extraction and the explain-ability problem of hidden layers; (3) establishing shared and specific-class deep dictionary learning method based on robust discrimination for face recognition. The shared deep dictionary learns the deep features of illumination, expression, age, and so on. The specific-class deep dictionary learns the robust discriminative features of face. Thus, it can improve the robustness of face recognition. This research will amplify the development of face recognition, enrich and develop the theoretical system of dictionary learning and deep learning.

为了增强人脸识别的鲁棒性，本项目研究如何利用深度学习和字典学习理论构造深度字典学习模型，从而提高字典提取人脸特征的性能。研究内容包括：(1)利用分解字典和合成字典构建基于自编码网络的深度字典学习模型，使字典模拟人脑的分层结构建立从底层信号到高层特征的映射，提取人脸的多层次深度特征，解决传统字典不能提取人脸深度特征的问题；(2)利用每层编码系数矩阵的行向量、分解系数和原子特征构造鲁棒判别式项模型，建立节点、权重和深度特征间的联系，使其不仅能继承上层特征的结构信息，又能保持该层特征的结构信息，增强深度字典提取特征的性能和隐藏层的可解释性；(3)构造基于共享和特定类的鲁棒判别式深度字典学习算法，使共享深度字典学习抵抗光照、表情和年龄等变化的深度特征，特定类深度字典学习人脸的鲁棒判别性深度特征，增强人脸识别的鲁棒性。本项目的开展将有利于人脸识别技术的发展，促进字典学习和深度学习理论的发展与完善。

传统字典学习属于浅层学习，不能提取人脸等数据的非线性深度特征。此外，判别式模型易受样本噪声影响且不能自适应更新，降低了它们在人脸识别等领域中的鲁棒性。. 在本项目中，申请人把profiles（编码或解码系数矩阵的行向量）引入到判别式模型中。利用profiles表示其对应原子重构(分解)样本集的贡献，提高字典学习的可解释性。其次，推导出profiles与原子的相似性定理，把基于样本和编码（分解）系数构造判别式模型方法转换为基于原子和profiles方法。该方法不受样本噪声影响且能自适应更新，增强判别式模型的鲁棒性和自适应性。在上述研究基础上，提出自适应局部约束、自适应局部保序约束和判别Fisher嵌入等鲁棒判别式模型，构造深度分解字典学习算法及字典学习和深度学习融合算法，设计基于它们的人脸识别算法。此外，还提出了基于PID控制器引导注意力卷积神经网络模型等深度学习算法，为构造深度字典学习模型提供理论和技术基础。. 本项目提出的概念、定理和算法为设计鲁棒判别式模型提供新的研究思路和方法，丰富了字典学习、深度学习和深度字典学习的相关理论体系，提升了它们在人脸识别领域中的性能，促进它们在图像识别和图像处理等领域中的应用。.