面向公共安全的不可控条件下异质人脸图像变换和识别研究

项目名称： 面向公共安全的不可控条件下异质人脸图像变换和识别研究

项目编号： No.61501339

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 王楠楠

作者单位： 西安电子科技大学

项目金额： 21万元

中文摘要： 在公共安全领域（例如刑侦破案或反恐）中，模拟画像和来自于视频监控中的图像是锁定犯罪嫌疑人的关键线索。本课题中，将来自于不同数据源的人脸图像（如模拟画像、视频监控图像和正常采集的人脸照片）统称为异质人脸图像。在实际应用中，这些异质人脸图像会被一起用来进行犯罪嫌疑人身份的排查与确认。由于模拟画像和视频监控图像与公共安全部门采集的照片数据分别在纹理和分辨率上存在很大差异，一般通过将它们先变换为同一模态进行识别，或者通过提取异质人脸图像间的共有表示信息进行识别。现有异质人脸图像变换和识别方法主要针对可控条件下获得的图像，而现实中获得图像往往是不可控条件下的（光照变化，背景复杂）。本项目拟通过多特征表示、概率图模型以及级联线性回归等思想实现不可控条件下异质人脸图像的变换和识别。本项目的研究能够辅助提高公共安全部门破案效率与准确率。

中文关键词： 数据表示；图像特征；图像分类；图像特征提取；图像信息

英文摘要： For public security (such as case solving and anti-terrorist), simulated sketches and images from video surveillance are important cues for suspect searching. In this project, face images from different imaging sources (e.g. simulated sketches, video surveillance images, and normal captured face photos) are called heterogeneous face images. In real-world applications, these heterogeneous face images are used together to check and identify suspects. The study found that conventional face recognition methods cannot be directly applied to heterogeneous face image recognition because simulated sketches and video surveillance images have great difference with images collected by departments of public security. In order to reduce these discrepancies, this project will transform these heterogeneous images into homogeneous images for identification, or extract their common representation for identification. Existing heterogeneous face image transformation and recognition methods are mainly proposed for images collected under well-controlled environments. However, most real-world images are captured in-the-wild with illumination changes or complicated background. This project will study the heterogeneous face image transformation and recognition problem in-the-wild by exploring the ideas of multi-view representation, probabilistic graphic models, and cascaded linear regression. This study would help related public security departments to improve the efficiency and accuracy of arresting suspects.

英文关键词： data representation;image feature ;image classification;image feature extraction;image information