毕业生知识库

大数据时代背景下海量人脸图像的获取已不是难事，丰富的训练数据给人脸图像的分析研究带来了机遇。然而互联网人脸数据含有大量噪声、缺少标注信息的特性也给人脸图像的分析研究带来了巨大的挑战。与此同时，大数据时代背景下对人脸数据处理的速度要求不断提高，如何快速高效而又准确地对人脸图像进行分析研究是当前人脸识别领域内的一大难题。本文从非监督学习和监督学习两方面入手，探索大规模人脸图像的聚类算法和识别算法，分别解决当前大量人脸数据无标注的问题和识别速度较慢的问题。具体研究内容如下：

• 提出了一种面向快速人脸聚类的贪婪子空间聚类算法。由于贪婪算法对初始状态的高敏感性，因此基于贪婪算法的子空间聚类算法的性能极大程度地受限于初始化状态。我们通过引入一个初始化子空间的构造方法，为贪婪聚类算法提供一个可靠的初始化子空间，进而确保贪婪聚类算法的效果。考虑到贪婪算法容易陷入局部极值的特点，我们将浮动搜索策略引入聚类算法，通过回溯步骤去除可能错误的邻接关系，进一步提升贪婪聚类算法的效果。本方法对噪声具有较好的鲁棒性，具备较好聚类效果的同时时间复杂度较低。符合大规模无标注人脸图像快速聚类的需求。

• 提出了一种基于定序神经网络的人脸识别方法，通过不同特征间的定序表示有效地减少网络参数。并使用三元组损失作为目标函数，明显提升人脸表示的泛化能力。在保证高准确率的同时，相比现有的神经网络模型的具备更少的网络参数，使得存储成本、计算成本大大降低，更加适应于大数据场景下的人脸图像处理任务。

总的来说，本文从人脸聚类和人脸识别两个方面分别提出了一种快速聚类算法和一种简约的神经网络模型，适用于互联网时代下大规模人脸图像的处理。

In the era of big data, it is not difficult any more to collect large number of face images. Abundant training data brings opportunity to the research of facial image analysis. However, face images crawled from Internet contain a lot of noise, and the lack of label information entails enormous challenge. Besides, with the promotion of face recognition technology, the speed requirement of face data processing improves a lot. How to efficiently and effectively perform the facial image analysis is a major problem in the field of face recognition.

The contributions of this thesis are summarized as follows:

• This thesis proposes a novel greedy subspace clustering strategy for face clustering. Greedy-based methods are sensitive to initialization for they are only stepoptimal. To alleviate the initialization problem of greedy-based method, a robust initialization step is induced. In order to control the complexity, the nearest subspace neighbor is added in a greedy way, and the subspace is updated by adding an orthogonal basis involved with the newly added data points in each iteration. A backtracking mechanism is introduced after each iteration to reject wrong neighbors selected in previous iterations. Our algorithm can significantly improve the clustering accuracy without sacrificing much computational time, which is suitable for efficiently face clustering.

• An ordinal-CNN is proposed which contains ordinal units. The ordinal-CNN can learn an ordinal representation of different features, which achieves variable selection and dimension reduction. Besides, the learned face representation has great generalization power for introducing a triplet loss. The proposed CNN models achieve state-of-the-art results int the LFW dataset. At the same time, a reduction of computational cost is reached, which makes our method applicable for large-scale facial image analysis.

In a word, this thesis studies large-scale facial image analysis via both unsupervised learning and supervised learning.