摘要
针对数据量巨大、类别多、真实类别数未知、样本数量不均衡、类内变化多的无标签人脸图像分类问题,提出基于附加间隔Softmax特征的近似等级排序人脸聚类算法。使用附加间隔Softmax损失结合Inception-ResNet-V1网络训练人脸识别模型来提取深度人脸特征,并应用于近似等级排序聚类。在LFW人脸数据集、LFW与视频模糊人脸的混合数据集上进行实验,结果表明该模型在人脸识别准确率、误识率为0.1%时的验证率均优于其他模型,近似等级排序聚类在F1度量得分优于其他聚类算法,具有更强的鲁棒性和应用价值。
In order to deal with the problem of unlabeled face classification with huge data,multiple categories,unknown number of real categories,unbalanced sample number and various changes intra class,this paper proposes an approximate rank order face clustering algorithm based on additive margin Softmax features.We used the additional margin Softmax loss combined with the Inception-ResNet-V1 network to train the face recognition model.Then,we used the model to extract the depth of face features and apply them to the approximate rank order clustering.Experiments were conducted on LFW face dataset and the mixed face dataset,LFW and video blurred faces.The experimental results show that our model is superior to other models in face recognition accuracy rate and verification rate when the error rate is 0.1%.And its approximate rank sorting clustering is better than other clustering algorithms in F1 measurement score.Therefore,this algorithm proposed in this paper has stronger robustness and application value.
作者
王锟朋
高兴宇
Wang Kunpeng;Gao Xingyu(University of Chinses Academy of Sciences,Beijing 100049,China;Institute of Microelectronics of the Chinese Academy of Sciences,Beijing 100029,China)
出处
《计算机应用与软件》
北大核心
2020年第2期111-117,共7页
Computer Applications and Software
基金
国家自然科学基金项目(61702491)。
