基于卷积神经网络的大规模人脸聚类

大规模人脸聚类不仅要求高效的人脸特征，还要求聚类算法在保持高准确率的同时耗时短.本文通过构建卷积神经网络高效提取人脸特征，并采用经典K-means算法和现阶段新颖的CFSFDP（Clustering by Fast Search and Find of Density Peaks）算法进行大规模人脸聚类.实验在聚类数目递增的情况下进行，并通过随机指标(Rand Index,RI)、信息熵、F₁-measure和混淆矩阵可视化来综合评估聚类的质量.结果表明，在大规模人脸聚类的情况下，卷积神经网络特征融合K-means的人脸聚类算法速度和准确率均优于CFSFDP算法.这一结论对大规模人脸聚类的实际应用具有重要的指导意义.

Large Scale Face Clustering Based on Convolutional Neural Network

The key challenge of large scale face clustering is to extract effective facial features and construct an accurate model with less time complexity. In this research, effective features are first extracted based on convolutional neural network (CNN). Then K-means, a classical cluster algorithm, and a state-of-art algorithm named CFSFDP (Clustering by Fast Search and Find of Density Peaks) are used to cluster large scale face images. Rand Index, entropy, F₁-measure and the visualization of confusion matrix are further applied to comprehensively assess clustering quality. All the tests are under the condition of the increasing numbers of clustering centers. Experiment results demonstrate that K-means has a better performance as well as a much higher speed than CFSFDP. This conclusion is believed to shed new light in the area of face clustering.