基于图像集人脸识别关键技术研究

Face recognition has traditionally been posed as the problem of identifying a face from a single image. With the increase of available video cameras and large capacity storage media, how to effectively obtain the accurate classification from image sets containing more information compared to a single image, i.e., matching a probe image set against all the gallery image sets each representing one subject, has attracted increasing interest in computer vision community. For the limitations of Image-set based facial feature learning of the current technology, this project aims to propose a fully automatic image set quality assessment, representation, and recognition approach，with using the deep learning techniques. The major topics include: (1) comprising facial image quality assessment and face frontalization based registration process, which normalize the complexly changed face sets to an unified form, to facilitate the following feature learning procedure; (2) introducing deep learning framework and sparse coding techniques into the image set representation which can automatically discover the underlying geometric structure; (3) designing adaptive distance metric learning technique which can exploit discriminative information for better characterizing the similarity between each gallery face and the probe image set. If this project will be carried out, we will obtain the improvements in the theory and techniques about face image set representation. Finally, we expect a new real-world face recognition system will be implemented based on the algorithms which will be deeply investigated in this project.

传统的人脸识别系统以单张人脸图像为研究对象，随着视频数据的广泛使用，如何利用多幅图像提供的信息获得更好的匹配性能以及如何对同一个体的人脸图像集合进行识别，已成为本领域新的研究热点。针对当前技术在人脸图像集特征表达上的不足，本项课题采用深度特征学习以及结构性稀疏编码方法的技术路线，提出一套面向实际应用环境的人脸图像集质量评估、建模与距离度量计算的新方法。本课题内容包括：（1）拟研究人脸图像集质量评价以及图像正面化处理方法，把包含复杂变化的人脸图像集归一化为统一的形式；（2）以深度神经网络学习以及结构性稀疏编码方法为技术点，研究鲁棒人脸图像集表达模型；（3）拟设计具有鉴别特性的自适应度量矩阵，增强同类图像集特征的判别性能，提高识别的泛化能力。本项研究成果将有助于解决真实环境下的人脸图像集识别难题，具有较大应用价值。

随着大数据时代的到来，利用视频数据进行准确的人脸识别成为当前的一大研究热点。传统的人脸识别方法适用场景较为单一，在复杂环境中往往难以提取具有足够判别性和表达力的人脸特征。针对当前技术在人脸图像集特征表达上的不足，本项课题采用深度模型、判别信息融合的技术路线，提出一套面向实际应用环境的人脸图像集质量评估、判别性表达与建模的新方法。本课题主要内容包括：（1）设计了人脸图像集图像质量评估及正面化重构算法，把复杂环境中采集的人脸图像转化为统一的形式；（2）设计具有鉴别特性的损失函数等，研究了人脸图像集判别特征表达方法；（3）以深度神经网络学习为主体结构，引入了迁移学习等技术点，研究鲁棒人脸图像集建模算法。项目研究算法成果已经应用于智能警务执法仪、个人法定证件采集等多个应用平台，取得了良好的效果。本项研究成果有助于解决真实环境下的人脸图像识别难题，具有较大应用价值。