深度卷积特征向量用于快速人脸图像检索-计算机辅助设计与图形学学报.PDF

30 12 Vol.30 No.12 2018 12 Journal of Computer-Aided Design Computer Graphics Dec. 2018 深度卷积特征向量用于快速人脸图像检索 李振东, 钟 勇, 曹冬平 ( 中国科学院成都计算机应用研究所研发中心 成都 610041) ( 中国科学院大学 北京 100049) (lizhendong13@) 摘 要: 针对计算机视觉领域的人脸图像检索计算复杂和检索时长问题, 提出一种基于深度卷积特征的快速人脸图 像检索算法. 首先使用人脸图像训练集对深度卷积神经网络模型进行人脸分类训练; 然后采用三元组损失方法对已 训练好的人脸分类网络模型进行微调, 更新网络的参数, 使得网络能够更加有效地提取不同人的人脸图像特征, 构 建高效的卷积特征向量进行人脸检索初步过滤; 为了进一步提高系统的检索性能, 提出一阶段查询扩展方法对待检 索人脸图像的卷积特征向量进行均值融合加强, 再次执行检索任务, 得到最终的检索结果. 在Celebrities Face Set 和 Labeled Faces in the Wild dataset 这2 个人脸检索数据集上进行实验的结果表明, 该算法不仅能够显著地提高检索结 果的准确率, 而且简单可靠, 能够快速地实现人脸检索任务. 关键词: 图像检索; 人脸检索; 卷积神经网络; 特征向量 中图法分类号: TP391.41 DOI: 10.3724/SP.J.1089.2018.17119 Deep Convolution Feature Vector for Fast Face Image Retrieval Li Zhendong, Zhong Yong, and Cao Dongping (Research and Development center, Chengdu Institute of Computer Application, Chinese Academy of Sciences, Chengdu 610041) (University of Chinese Academy of Sciences, Beijing 100049) Abstract: In order to solve the problem of complexity in calculating and long time in retrieving of face im- age retrieval in the field of computer vision, a face image retrieval algorithm based on deep convolutional features is proposed. Firstly, the convolutional neural network model is trained for face classification by face image training data set. Then, the trained face classification network model is fine-tuned by using the triplet loss method, and the parameters of the network are updated to make the network more efficiently in extract- ing face image features of different people and construct efficient convolution feature vectors for preliminary filtering stage of face retrieval. In order to further improve the