基于深度卷积神经网络与中心损失的人脸识别

传统人脸识别方法手工设计特征过程复杂、识别率较低,对于开集人脸识别通用深度学习分类模型特征判别能力较弱。针对这两方面的不足,提出了一种以分类损失与中心损失相结合作为模型训练监督信号的深度卷积神经网络。首先,利用构建的应用场景数据集优调从公共数据集获得初始化参数的深度人脸识别模型,解决训练数据过小和数据分布差异问题,同时提高模型训练速度;然后,以传统损失函数和新的中心损失作为迁移学习过程中的监督信号,使得类内聚合、类间分散,提高模型输出人脸特征的判别能力;最后,对人脸特征进行主成分分析,进一步去除冗余特征,降低特征复杂度,提高人脸识别准确率。实验结果表明,与传统人脸识别算法相比该算法可以自动进行特征提取,并且相对于通用深度学习分类模型该算法通过度量学习使特征表示更具判别力。在自建测试集和LFW、YouTube Faces标准测试集上都取得了较高的识别率。

Face recognition based on deep convolution neural network and center loss

For traditional face recognition methods,the process of manual design features is complex and face recognition rate is low. The feature discrimination ability of general deep learning classification model is weak, for the open set face recognition. Aiming at these two problems, a kind of deep convolution neural network is proposed, which combines the classification loss with the central loss as the model training monitoring signal. Firstly, a deep face recognition model based on the initialization parameters obtained from the public dataset is fine tuned using application scene dataset,which can effectively solve the problem of training data is too small and data distribution differences and improve the training speed of the model. Then, the traditional loss function and the new central loss are used as the monitoring signals in the process of transfer learning, which can make the intra-class aggregation and inter-class dispension and improve the discriminative ability of the model output features. Finally, the principal component analysis is used to remove the redundant face features, reduce the complexity and improve face recognition rate. The experimental results show that our algorithm can automatically extract features compared with the traditional face recognition algorithm and relative to the general deep learning classification model, the algorithm makes the feature representation more discriminative with metic learning. A higher recognition rate has been achieved in the self built test set and the LFW and YouTube Faces Standard test sets.