基于表达残差稀疏性的遮挡人脸识别方法研究

Face recognition technology has become one of the hottest research topics in image processing and pattern recognition in the latest three decades. Nowadays, we are facing a great challenge on how to make the authentication with occluded face. Conventional methods seek to increase the robustness of classifiers to occlusions. But these means cannot meet the requirement of high accuracy as well as lack of strong theoretical support. This project uses the theory of linear coding to solve the problem of occluded face recognition. We hold that the residual between test samples (with occlusions) and its prediction represented by a combination of the training samples is sparse. Based on such property, we put forward three novel methods with different features. The first one is a rapid classification method which uses training samples of each class to represent a test sample. The second one owns higher recognition accuracy via collaborative representation. The third one has higher robustness by introducing nonlinear function. In our opinion, this project has certain theoretical innovative value, because we introduce a new theory to solve problem of occluded face recognition. We believe that our project can promote the development of face recognition technology and be helpful to the construction of information society and safe city.

人脸识别技术成为近三十年里图像处理和模式识别中最热门的研究主题之一。如何识别被遮挡的人脸是该技术目前面临的一个重大挑战。传统的方法通过提高分类器的鲁棒性来降低遮挡对识别结果的影响，然而效果不够理想且理论基础不够坚实。本项目提出利用线性编码理论来解决遮挡人脸识别问题。我们认为测试样本（带遮挡人脸图片）与用训练样本所表达的测试样本之间的残差满足稀疏性。基于这个性质，我们拟提出三种具有不同特性的新分类方法：第一种方法采用每类样本分别对测试样本进行表达的策略进行快速分类；第二种方法采用共同表达策略来进行更高准确性的分类；第三种方法采用引入非线性函数的策略进行具有更高鲁棒性的分类。由于我们利用了一种全新的理论所构建的方法来解决遮挡人脸识别问题，所以本项目具有一定的理论创新价值。通过本项目的研究，将能进一步推动人脸识别技术的发展，并对我国正大力倡导的信息化社会建设和平安城市建设等起到一定的推动作用。

在人脸识别系统中，待识别人脸上可能会带有各种不可预计的干扰因素，本研究主要关注如何识别部分被遮挡的人脸。本研究的理论基础是线性编码理论，它具有对测试样本被稀疏噪声干扰时仍然进行正确编码的特性。主要研究内容包括：基于单类别线性模型的分类方法；利用联合表达技术现实更高精确度的分类方法；利用鲁棒线性函数构建适用范围更广的分类方法。本研究发表人脸识别相关的SCI收录论文六篇（其中二区三篇，三区一篇，四区两篇），Google学术显示这六篇引用次数超过80次，申请人脸识别专利一项，超过项目预期成果数量要求。项目还支持其它SCI论文发表两篇。所发表的成果符合项目预期的技术路线，证明了项目设计的合理性和有效性。通过开展本项目的研究工作，取得一定的成果，其丰富了人脸识别领域的理论基础，推动了本领域的发展。