针对计算机视觉领域的人脸图像检索问题,提出了一种基于深度特征的快速人脸图像检索方法。该方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;在此基础上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,使得网络能够更加有效地提取人脸特征构建高效的特征向量进行人脸检索初步过滤;最后,为了进一步提高系统检索性能,提出一阶段查询扩展方法对待检索人脸图像特征向量进行融合加强。在两个公用人脸数据集(CASIA-3D FaceV1和Labeled Faces in the Wild dataset)上进行详尽的实验验证,结果表明,基于深度特征的人脸图像检索方法不仅能够显著提高检索结果的准确率,而且该方法简单可靠,能够快速地实现人脸检索任务。

In order to solve the problem of face image retrieval in the field of computer vision, a face image retrieval method based on the deep features is proposed. Firstly, the convolutional neural network model is trained for face classification by face image training data set. Based on this, the triplet loss method is used to fine-tuning the trained face classification network model so that the network can be more efficient to extract face features of different people and construct efficient feature vectors for preliminary face retrieval filtering stage. In order to further improve the performance of system retrieval, the one-stage query expansion method is proposed to reconstruct the eigenvectors of face images to be retrieved. Through exhaustive experimental verification on two public face datasets (CASIA-3D FaceV1 and Labeled Faces in the Wild dataset), the results show that the face image retrieval method based on deep features improves the accuracy of the retrieval results significantly. Moreover, this method is simple and reliable, and can quickly realize the task of face retrieval.