大光照变化单样本人脸类内变化建模研究

In outdoor visual monitoring system, the face images are with a wide range of holistic and local illumination variations. The single sample face intra-class variation modeling under wide illumination variations is a tough problem of face image processing and recognition area in the visual monitoring system. This project will focus on researching two key scientific issues the wide illumination variation processing and the single sample face intra-class variation modeling. On the one hand, this project will research to construct the wide illumination variation control model based on spatial feature decomposition, which can transform the wide illumination variations into the narrow range smooth illumination, based on above model, the multi-local-neighborhood based illumination invariant measure model with dynamic compensation will be established to eliminate the measure distortion and enhance the face discriminant features. On the other hand, this project will research the face linear intra-class variation model with removing redundant information, further the face nonlinear intra-class variation correction model based on deep auto-encoder will be established to eliminate face nonlinear intra-class variation, and then the robust face intra-class variation model can be formed. The research results of this project will form a novel method to solve the single sample face intra-class variation modeling under wide illumination variations. This project will be of great theoretical and practical value to enrich the research content of face image processing and recognition in the visual monitoring system, and promote the application of single sample face recognition in the area of public security surveillance.

在户外视觉监控系统中，人脸图像存在大范围全局和局部光照变化。大光照变化单样本人脸类内变化建模是视觉监控系统人脸图像处理和识别领域的难点问题。本项目将围绕大光照变化处理和单样本人脸类内变化建模两个关键科学问题展开研究，一方面，研究构建基于空域特征分解的大光照变化控制模型，将大光照变化转化为范围较窄的平缓光照变化，在此基础上，研究构建多局部邻域动态补偿的光照不变测度模型，消除测度失真、增强人脸鉴别特征。另一方面，研究构建消除冗余信息的人脸线性类内变化表示模型，进一步，研究构建基于深度自编码器的人脸非线性类内变化校正模型，消除人脸非线性类内变化，形成鲁棒人脸类内变化模型。本项目的研究将形成解决大光照变化单样本人脸类内变化建模的新方法，对丰富视觉监控系统人脸图像处理和识别的研究内涵，推动单样本人脸识别在公共安全领域的应用具有重要的理论意义和应用价值。

在户外视觉监控系统中，人脸图像存在大范围全局和局部光照变化。大光照变化单样本人脸类内变化建模是视觉监控系统人脸图像处理和识别领域的难点问题。本项目围绕大光照变化处理和单样本人脸类内变化建模两个关键科学问题展开研究，具体是，1）构建了基于空域特征分解的大光照变化控制模型，将大光照变化转化为范围较窄的平缓光照变化；2）构建多局部邻域动态补偿的光照不变测度模型，消除测度失真、增强人脸鉴别特征；3）构建消除冗余信息的人脸线性类内变化表示模型，以及基于深度学习的人脸非线性类内变化模型，形成鲁棒人脸类内变化模型。本项目的研究形成了解决大光照变化单样本人脸类内变化建模的新方法。.在项目资助下，共发表科研论文17篇，其中SCI期刊10篇，EI期刊和会议7篇，论文发表在IEEE T-IP，T-ITS，CVPR和PR等权威期刊和会议上。申请国家发明专利8项。获2020年中国发明协会发明创业创新一等奖(4/6)。联合培养2名博士研究生3名硕士研究生均已毕业，其中获2021年江苏省优秀硕士学位论文1篇。.项目研究的大范围光照变化处理算法可应用于大场景夜晚监控图像增强，使得被黑色覆盖的夜晚监控图像内容能够被监控人员辨识，从而实现夜晚大场景场面的有效监控。该算法已经在安徽省池州市九华山机场夜晚机场场面监控平台进行测试。