人脸认证作为人脸识别的一个分支，在维护社会稳定及个人安全方面具有重要的意义。目前，人脸认证方法主要有：以支持向量机、k最近邻、决策树等浅层模型为代表的传统的模式识别方法和以卷积神经网络为代表的深度学习方法。传统模式识别方法的识别精度依赖人工提取特征的结果，易受外界条件影响，智能化程度低。深度学习方法以自主学习方式能从复杂的数据中提取人脸的隐性特征，且泛化能力强。为此，本文提出两种基于深度学习的人脸认证算法。具体工作如下：

(1) 针对现有的基于深度学习人脸认证方法数据标记成本大，在训练样本较少的数据集上模型训练效果不佳的问题，提出了融合LeNet-5和Siamese神经网络的人脸认证算法。首先，将人脸数据匹配为成对样本并送入网络，采用Siamese神经网络框架，构建双分支LeNet-5卷积网络进行人脸特征提取，通过缩小卷积核、增加卷积层、改变激活函数调整模型结构；然后，使用Contrastive Loss函数进行网络优化，使类内差异最小化，类间差异最大化，提升网络对样本的区分能力；最后，通过度量样本特征间相似性判断样本类别。该方法将深度学习和度量学习结合起来，既避免了复杂的人工特征提取，又简化了人脸认证的流程。实验结果表明，该方法在训练样本较少的人脸数据集上的识别精度有明显提高。

    (2) 针对融合LeNet-5和Siamese网络在人脸表情以及光照等发生巨变的人脸数据集上特征表达能力不强、识别准确率不高的问题，提出了基于残差结构的融合LeNet-Residual与Siamese网络的人脸认证算法。首先，在数据匹配之前对图像进行局部纹理特征增强预处理，通过伽马校正、DoG滤波、对比度均衡化等操作消除过度曝光或阴影对图像产生的不良影响；然后，在融合网络的基础上引入A、B两种残差单元，设计双分支LeNet-Residual卷积神经网络提取更丰富的人脸特征，使用Contrastive Loss进行网络优化；最后，度量样本特征间相似性判断样本类别。实验结果表明，该方法在人脸表情、光照等发生较大变化的人脸数据集上特征表达能力更强、认证分类精度更高。

Face verification, as a branch of face recognition, is of great significance in maintaining social stability and personal safety. At present, face verification methods mainly include: traditional pattern recognition methods represented by shallow models such as support vector machines, k-nearest neighbors, and decision trees, and deep learning methods represented by convolutional neural networks. The recognition accuracy of traditional pattern recognition methods depends on the results of manual extraction of features, which is easily affected by external conditions and has a low degree of intelligence. Deep learning methods can extract the hidden features of human faces from complex data in an autonomous learning manner, and have strong generalization ability. Therefore, this paper proposes two face verification algorithms based on deep learning. The specific work is as follows:

(1) Aiming at the problem that the existing deep learning face verification method has high data labeling cost and poor model training effect on the data set with few training samples, a face verification algorithm combining LeNet-5 and Siamese neural network is proposed. First, the face data is matched as a pair of samples and sent to the network. The Siamese neural network framework is used to construct a two-branch LeNet-5 convolution network for face feature extraction. By reducing the convolution kernel, increasing the convolution layers, and changing the activation function adjusts the model structure; Then, the contrastive loss function is used for network optimization to minimize intra-class differences and maximize inter-class differences, and improve the network's ability to distinguish samples; Finally, the sample category is judged by measuring the similarity between sample features. This method combines deep learning and metric learning, which not only avoids complex artificial feature extraction, but also simplifies the face verification process. Experimental results show that the recognition accuracy of this method on face data sets with fewer training samples is significantly improved.

(2) In order to solve the problem that the feature expression ability and recognition accuracy of the fusion LeNet-5 and Siamese networks are not strong in the face data sets with great changes in facial expression and illumination, a fusion of LeNet-Residual and Siamese based on residual structure is proposed. First, before the data matching, the images are preprocessed with local texture feature enhancement, and the adverse effects of overexposure or shadow on the image are removed through gamma correction, DoG filtering and contrast equalization; Then, it is introduced on the basis of the fusion network two residual units, A and B, design a two-branch LeNet-Residual convolutional neural network to extract richer facial features, and use contrastive loss to optimize the network; Finally, measure the similarity between sample features to determine the sample category. The experimental results show that the method has stronger feature expression ability and higher classification accuracy on the face data set with large changes in facial expressions and lighting.