基于深度学习的大规模人脸图像检索

摘要

针对传统人工设计特征鲁棒性差、检索计算复杂等缺点,提出以深度特征表示人脸图像并采取由粗到细的人脸图像检索方法.首先,使用拥有近四百万张人脸图片的数据库训练卷积神经网络得到人脸特征提取模型;然后进行人脸特征提取、存储和聚类分析;最后,采用由粗到细地检索方式进行人脸检索.在LFW数据库上进行验证,基于深度学习的人脸特征的人脸检索准确率为99.1％,人脸检索时间约0.5s.实验结果表明,基于深度学习的人脸特征鲁棒性强、检索计算复杂度低.由粗到细的检索方法效率高,结果准确率高.%With the improvement of hardware,deep learning features overcome the weakness of traditional man-made feature such as poor robustness and complex retrieval calculation.A coarse-to-fine face image retrieval based on deep learning feature was proposed.First,a face feature extraction model is developed by using nearly four million face images to train the convolutional neural networks.Second,face feature is extracted,stored and clustered.Finally,face retrieval is performed by coarse-to-fine retrieval.Face verification gets a 99.1％ accuracy via deep learning face feature in the LFW benchmark and face retrieval costs only about 0.5 second in a million face retrieval benchmark.The experiment results illustrate that deep learning face feature is more robust and lower in computational complexity.The retrieval method from coarse to fine has high efficiency and high accuracy.