Journal of Frontiers of Computer Science and Technology ›› 2019, Vol. 13 ›› Issue (7): 1195-1205.DOI: 10.3778/j.issn.1673-9418.1812016

Previous Articles     Next Articles

Densely Connected Convolutional Networks Face Recognition Algorithm Based on Weighted Feature Fusion

WANG Xiaoyu+, HAN Changlin, HU Xinhao   

  1. School of Computer Science and Technology, Harbin University of Science and Technology, Harbin 150080, China
  • Online:2019-07-01 Published:2019-07-08

加权特征融合的密集连接网络人脸识别算法

王小玉+,韩昌林胡鑫豪   

  1. 哈尔滨理工大学 计算机科学与技术学院,哈尔滨 150080

Abstract: In unconstraint condition, face recognition is usually affected by expression change, angle deviation, occlusion and exposure in different degrees, also other comprehensive factors. At the same time, interfering factors exist in deep convolutional neural networks, such as lots of parameters, gradient diffusion or disappearance during training. To solve mentioned concerns, this paper proposes a FuseNet network model, which effectively utilizes local feature information such as eyes, nose and mouth as well as global feature information like facial contour, and puts forward the multi-loss function to reduce the gap between the class features and increase further distance between class characteristics, which effectively enhances the robustness of face recognition under constraint conditions. The weighted, densely connected convolutional neural network, is adapted to extract the global features of faces in which densely connected module effectively solves the problem like parameter redundancy and gradient diffusion caused by the deep network. Additionally, proposed different connection weights enable the network to make full use of selected characteristics. The experimental results show that the proposed FuseNet network achieves better recognition rate and robustness, not only in the closed CASIA-WebFace dataset, but also in open FLW dataset and MegaFace dataset.

Key words: face recognition, weighted densely connected, weighted feature fusion, multi-loss function

摘要: 在非约束条件下人脸识别常受到表情变化、视角偏差、不同程度的遮挡和曝光等各种综合因素的影响;并且深度卷积神经网络几乎都存在参数过多,训练时梯度扩散或消失等问题。针对上述问题,提出了FuseNet网络模型。该模型有效地利用了人的眼睛、鼻子、嘴巴等局部特征信息,同时又包含面部轮廓等全局特征信息,并提出了多损失函数进一步缩小类内特征差距和扩大类间特征距离,有效地增强了非约束条件下人脸识别的鲁棒性。通过使用加权密集连接卷积神经网络来提取人脸的全局特征,密集连接模块可有效地解决深层网络所引起的参数冗余以及梯度扩散等问题。不同的连接权值使得网络能够充分地利用各部分特征。实验结果表明,无论是在闭集的CASIA-WebFace数据集上,还是开集的FLW数据集、MegaFace数据集上,提出的FuseNet网络都具有较好的识别率和泛化能力。

关键词: 人脸识别, 加权密集连接, 加权特征融合, 多损失函数