【摘要】 当前基于深度卷积神经网络的人脸表示学习方法需要利用海量的有标注的人脸数据。在实际应用中,精确标注人脸的身份非常困难。因此,提出了一种基于前向无监督卷积神经网络的人脸表示学习方法。其中,基于Kmeans聚类获取训练样本虚拟标签,再利用线性判别分析进行卷积核学习。提出的网络结构简单有效,训练阶段不需要反向传递,训练速度显著优于有监督的深度卷积神经网络。实验结果表明,提出的方法在真实条件下的人脸数据集LFW和经典的Feret数据集上取得了优于当前主流的无监督特征学习方法和局部特征描述子的性能。更多还原

【Abstract】 The existing face representation learning methods based on deep convolutional neural network demand massive labeled face dataset.In real-world application,it is difficult to precistly annotate the labels of face dataset.In this paper,an unsupervised forward convolutional neural network based face representation learning algorithm was proposed.By design,virtual labels of training samples were abtained based on K-means clustering and then convolution kernels were learnt by linear discriminant analysis.The network architecture in this paper is simple and effecitve and does not need back propagation during training,so its training speed is much quicker than supervised deep convolution neural network.The experimental results demonstrate that the proposed method in this paper out performs the state-of-the-art unsupervised feature learning algorithm and local feature descriptors in both real-world Labeled Face in the Wild(LFW)dataset and the classical controlled Feret dataset.更多还原