基于深度网络的多形态人脸识别

Face Recognition with Multiple Variations Using Deep Networks

在实际的自动人脸识别系统中,输入的识别图像往往在表情、分辨率大小以及姿态方面呈现出多种变化。现在很多方法尝试通过线性或局部线性的映射来寻找由这些变化共享的统一的特征空间。利用由受限玻尔兹曼机(RBM)堆叠成的深度神经网络来发掘这些变化内在的非线性表达。深度网络能够学习高维数据到低维数据的映射关系,并且有助于提高图像分类和识别的性能。同时,为了实现在一个统一的深度框架下同时进行特征提取和识别,在网络的顶层增加了一个监督的回归层。在预训练阶段,通过训练集中不同姿态、不同表情以及不同分辨率的图像对网络进行初始化。在微调阶段,通过网络的输出与标签之间的差并利用标准反向传播的方法对模型的参数空间进行调整。在测试阶段,从测试库中随机选择一幅图像,获得统一空间下的特征向量。通过与参考图像库中的所有特征向量进行对比,利用最近邻域的方法识别人脸身份。在具有丰富表情以及大姿态变化的CMU-PIE人脸数据库上进行了全面的实验,结果表明,提出的方法取得了比最新的局域线性映射(或局部线性)的人脸识别方法更高的识别率。

In automatic face recognition（AFR） applications, input images typically present multiple types of variations on expression, resolution and pose. Existing approaches attempt to seek a common feature space shared by these varia- tions through linear or local linear mappings. We used deep networks stacked by restricted Boltzmann machines to dis- cover intrinsic non-linear representations of these variations. Deep learning can provide insight into how high-dimension- al data are organized in a lower dimensional feature space and it also improves the performance of classification and rec- ognition. In the meantime, we realized a supervised regression layer on the top of the network so that both feature ex- traction and recognition can be achieved in a unified deep framework. For the pre-training phrase, the whole network is initialized by training set including different poses with various expressions under high resolution（HR） and low resolu- tion（LR）. For the fine-tuning phrase, the parameter space is adjusted by the errors between the output of network and the labels via standard back propagation. For the test phrase, a profile face image from Probe is chosen randomly, then the feature vector in the subspace is gained. Compared with all of the vectors in the Gallery set, we determined the iden- tity of images by the nearest neighborhood. We performed the extensive experiments on CMU-PIE facial database that presents rich expressions and wide range pose variations. The experiments show the superior recognition rate of our ap- proach over the state-of-the-art linear（or locally linear） methods_