神经网络多层特征信息融合的人脸识别方法

由于人脸面部结构复杂，不同人脸之间结构特征相似，导致难以提取到十分适合用于分类的人脸特征，虽然神经网络具有良好效果，并且有很多改进的损失函数能够帮助提取需要的特征，但是单一的深度特征没有充分利用多层特征之间的互补性，针对这些问题提出了一种基于神经网络多层特征信息融合的人脸识别方法。首先选择ResNet网络结构进行改进，提取神经网络中的多层特征，然后将多层特征映射到子空间，在各自子空间内通过定义的中心变量进行自适应加权融合；为进一步提升效果，将所有特征送入Softmax分类器，同时对分类结果通过相同方式进行自适应加权决策融合；训练网络学习适合的中心变量，应用中心变量计算加权融合相似度。在同样的有限条件下，在使用AM-Softmax损失函数的基础上，融合特征在LFW(Labeled Faces in the Wild)上的识别效果了提升1.6%，使用融合相似度提升了2.2%。能够有效地提升人脸识别率，提取更合适的人脸特征。

Face recognition method based on neural network multi-layer feature information fusion

Because the structure of the face is complex and the structural features of different faces are similar, it is difficult to extract facial features that are suitable for classification. Neural networks generate good results, and the recent improvements in many loss functions can help extract the required features. However, a single depth feature does not make full use of the complementarity of multi-layer features. To solve these problems, we propose a face recognition method based on the fusion of neural-network multi-layer feature information. First, we select the ResNet network structure to improve the outcome, then we extract the multi-layer features in the neural network. These features are then mapped onto the sub-spaces. Next, adaptive weighted fusion is performed of the defined central variables in the respective sub-spaces. To realize further improvement, all the features are sent to the Softmax classifier, and the classification results are fused in the same way by adaptive weighted decision-making. The training network learns the appropriate central variable, which is applied to calculate the weighted fusion similarity. Under the same conditions, based on the AM-Softmax loss function, the recognition of the fusion feature on the Labeled Faces in the Wild database increased by 1.6%, and the fusion similarity increased by 2.2%. We conclude that the proposed method effectively improves the face recognition rate and extracts more suitable facial features.