基于人脸关键点与增量聚类的多姿态人脸识别

Multi-Pose Face Recognition Based on Facial Landmarks and Incremental Clustering

人脸姿态变化复杂且对人脸识别性能影响明显,提出了一种融合LCCDN (LSTM and CNN based Cascade Deep Network)与增量聚类的多姿态人脸识别方法。采用LCCDN模型定位人脸关键点,利用长短时记忆网络(LSTM)的记忆功能寻找人脸各关键点在空间上的全局上下文的依赖关系对人脸关键点初始化,并通过卷积神经网络模型,采用由粗到精的策略;定位人脸关键点;以人脸关键点作为人脸朝向描述子,同时为适应人脸姿态不断地动态更新,采用基于熵诱导度量机制的增量聚类方法,对头部姿态进行动态增量聚类,构建人脸姿态池。在此基础上,通过建立不同姿态的人脸识别分类模型实现多姿态人脸识别,在CAS-PEAL-R1、CFP和Multi-PIE三个数据集上的人脸识别准确率分别达到96.75%,96.50%,97.82%。通过与同类人脸识别方法的客观定量对比,实验结果表明所提方法有效、可行。

Owing to complex changes in face pose and the obvious influence on face recognition performance, a new approach is proposed for multi-pose face recognition based on the fusion of the LSTM (long short term memory network) and convolutional neural network-based cascade deep network (LCCDN ) and incremental clustering. First, a LCCDN is designed to locate facial landmarks, and the memory function of the LSTM in LCCDN is used to explore the spatial contextual information between facial landmarks;then, facial landmarks are initialized. ACNN network model is used to fine facial landmarks by employing a coarse-to-fine strategy. Next, we consider the facial landmarks as face orientation descriptors. Simultaneously, to adapt to the dynamic updating of the face pose, an entropy-induced metric-based incremental clustering method is used to construct a face-pose pool by dynamically clustering head poses. In this manner, multi-pose face recognition is realized by establishing various face classification models with different poses. The recognition accuracies using the CAS-PEAL-R1, CFP, and Multi- PIE datasets are 96.75%, 96.50%, and 97.82%, respectively. In addition, comparisons with existing multi-pose face recognition methods highlight the superior performance of the proposed method.