基于多特征融合的素描人脸识别研究

One of the most difficult challenges in automated face recognition is computing facial similarities between face images acquired in alternate modalities. Called heterogeneous face recognition (HFR), successful solutions to this recognition paradigm will allow the vast collection of face photographs (acquired from mug shots and other sources of frontal face images) to be matched against face images from alternate modalities(e.g. forensic sketches, infrared images).Sketch face recognition, which belongs to HFR, are studied. This project propose some ideas based multi-features fusion(e.g. LBP, SIFT, Gabor, and other Key regional feature) to provide some clues on how the HFR is better solved. The contents include: Extract topological consistency feature; Compute the distance between the topological consistency feature by using manifold space; Extract effective local feature; Build sketch face database and analyze algorithm time complexity. The technology route is as follows: Firstly, face feature are classified based on Stepwise Refinement thought.Secondly, the feature are blocked adaptively according to the number of the key points in the image pixel neighborhood area. Lastly, features are selected to distinguish by cascade Adaboost. Furthermore, the objective function is built to extract the topological consistency feature by using vector distance between the key points descriptors, phase difference and gradient information. The topological consistency feature includes texture information and structure information. all images of the same subject are located on the same manifold surface, and the distance between within-class or between-class is computed reasonably on this surface or between two surfaces, which suggests the difference in the same subject or the different subjects. The dimension of the topological consistency feature is reduced by using manifold learning theories. In summary, the project may help us understand the discipline of human cognition, and has an important scientific significance.

素描人脸识别属于异质人脸识别分支，是指把来自大头照的图像与刑侦素描等不同模态的人脸相匹配，已成为人脸识别最困难的挑战之一。本项目把用于刑侦的素描人脸识别作为研究对象，把LBP、SIFT、纹理走向等多特征融合起来，试图为解决上述问题提供方案。内容包括：.提取拓扑一致性特征；利用流形空间计算拓扑一致性纹理特征之间的距离；提取有效的分块局部特征；构建素描人脸库并分析算法性能。技术路线为：基于逐步求精的思想对人脸特征进行分类，根据邻域内关键点数量进行特征自适应分块，利用级联AdaBoost分类器挑选特征进行识别；此外，利用关键点描述子之间的向量距离、相位变化和梯度信息，构建目标函数，提取拓扑一致性特征，该特征包含了图像的纹理信息和结构信息，然后利用流形学习思想对拓扑一致性特征进行降维，并把类内或类间之间的差别看做流形空间中曲面内部或曲面之间的距离。本项目可进一步了解人类的认知规律，科学意义重大。

素描人脸识别是指把来自大头照的可见光图像与素描人脸图像相匹配，由于可见光图像和素描图像存在较大的模态差异、素描图像样本偏少，因而它已成为人脸识别领域最困难的挑战之一。本项目以人脸识别理论作为指导，把用于刑侦的素描人脸识别作为研究对象，从特征提取、度量距离、素描人脸合成、超分辨率重建等多个角度出发，试图为解决上述问题提供方案。主要研究提取拓扑一致性特征、利用流形空间计算拓扑一致性纹理特征之间的距离、提取有效的分块局部特征、基于最优相关的素描人脸合成算法、构建素描人脸库并分析算法性能以及超分辨率重建。创新性的成果如下：. （1）针对传统特征缺少关键信息的问题，提出了一套新的针对素描人脸识别的自适应分块与权重计算方法——多尺度融合与权重分配,一种在小样本空间中基于高斯过程的快速人脸验证方法和一种基于深度迁移学习的素描人脸识别框架。. （2）针对异质人脸模态间距大的问题，提出了一种基于深度残差网络和度量学习的素描人脸识别模型和一种基于对偶马氏损失的素描人脸识别方法。. （3）提出了一种基于自适应分块和Gabor特征提取相结合的素描人脸图像合成方法和一种感知哈希算法（perceptual hash，简称pHash）与稀疏编码（sparse coding，简称SC）相结合的素描人脸合成方法。. （4）提出了一种基于双层生成对抗网络的素描人脸合成方法，一种基于深度学习的素描人脸合成方法，一种基于三网络对抗学习的高质量素描人脸合成方法以及一种基于多判别器循环生成对抗网络的素描人脸合成方法。. （5）提出了一种基于变形判别自动编码器的人脸超分辨率算法和一种基于双层级联神经网络的人脸超分辨率重建模型。. 此外，本项目成果与其他算法融合可应用于智能交通信号处理领域，在雷达、视频等多传感器数据处理与融合方向，发挥重要作用，可有效提高目标识别与跟踪精度。相关成果获北京市科学技术奖和中国智能交通协会科学技术奖，发表中文核心期刊论文16篇，SCI检索期刊论文8篇，EI检索论文5篇，出版学术专著1部，授权发明专利4项，登记计算机软件著作权12项，科技成果转化1项，举办国际会议1次，培养硕士研究生14人，其中3人获校级优秀硕士学位论文。