CASIA OpenIR  > 毕业生  > 博士学位论文
基于光照分析的人脸图像取证算法研究
张旭
Subtype博士
Thesis Advisor彭思龙
2019-06-03
Degree Grantor中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword图像取证 光照分析 人脸图像 光照估计 透视投影 活体检测
Abstract

随着相机、手机等拍摄工具以及互联网的普及,以及影像编辑工具和处理技术的不断发展,普通群众也可以方便的修改影像,在给人们生活带来便利的同时,也使得人们对所见影像的真实性不再深信不疑。特别是将一个人的头像放置在某张照片或某段视频中,制造该人参与该影像所呈现的事件的假象,是最常见也是危害性最大的一类恶意篡改。而近期由于网约车客户端人脸验证以及支付宝刷脸支付被人脸照片/视频攻击而造成的恶劣事件,同样使得同人脸相关的安全问题受到重视。研究人脸图像真实性的取证技术具有重大意义。
已有方法从数字真实性、内容真实性和对象真实性等方面开展图像取证的研究,在取得相关效果的同时,也有各自的使用条件和适用范围,并容易被反取证手段攻击。实际应用中,经常需要多种取证方法配合才能给出有效的鉴定意见。作为成像的基本要素,光照虽然得到了广泛研究,并在图像取证中得到应用,但是仍然有很多值得深入探索、可以完善提高、可以灵活应用的地方。因此,本文将研究目标定位为基于光照分析对人脸图像取证方法进行研究,得到准确性更高、鲁棒性更好、说服力更强的取证方法,并在实际问题中进行应用。本文主要研究内容及创新点有以下3点:

1.人脸上复杂阴影、重建人脸模型的误差会严重影响现有光照估计方法的准确性。本文利用阴影客观反映环境光照的物理事实,提出了基于环境光/漫反射光分离的光照估计方法(Light Estimation based on Ambient/Diffuse light separationLEAD)。首先提出一种更符合物理事实的环境光/漫反射光分离的球面谐波光照模型(Sphere Harmonic lighting based on Ambient/Diffuse light separationSHAD)。在此基础上,利用迭代的方法估计环境光照,进而估计漫反射光照。实验表明,相比于现有的光照估计算法,本文方法增强了光照估计对于毛发纹理、复杂阴影、以及法向量误差等因素的鲁棒性。利用该方法得到的光照,结合环境光等信息进行光照一致性判断,可以得到更加符合客观实际、更令人信服的判断意见。

2.相机模型是透视模型,而传统方法则是基于平行投影和弱透视投影来估计光照,这样直接比较进行一致性分析,就会有本质上的误差。本文提出了一种透视投影下物体的空间光照估计算法。通过将各物体坐标系统一到相机坐标系下,得到各物体相对于相机坐标系的空间光照,从物理原理上使得光照估计方法的准确度得到了显著提高。实验表明,相比于传统方法基于平行投影光照进行光照一致性分析,采用本文方法得到的空间光照进行光照一致性分析具有更高的准确度,结合等效焦距、人脸空间位置及重新透视投影图像等相关信息进行照片空间合理性分析,得到的判断意见具有更强的说服力。

3.根据照片和视频中的物体固定在照片和屏幕中、无法对场景光照变化做出响应,而真实人脸光照分布则随场景光照变化而变化的事实,本文提出了一种基于光照变化分析的活体检测方法,通过分析场景光照改变前后采集的人脸光照变化同场景光照变化的相关性来区分真实人脸和欺骗人脸。经过光照估计、改变光照、计算相关性及假设检验判断等步骤,可以有效进行活体判断。实验结果表明,本文提出的活体检测方法对多种形式的欺骗人脸都可以有效检测,并且具有无需用户交互、使用条件宽松、应用范围广、设备要求简单、检测原理不易被察觉、有助于人脸识别等优点。

Other Abstract

With the popularity of cameras, mobile phones and the Internet, as well as the continuous development of video editing tools and video processing technologies, people can easily modify the images, which brings convenience to life. But it also raises doubts about the authenticity of the images. In particular, placing a person’s face into another photo or a piece of video, creating the illusion that the person is involved in the event presented by the image, is the most common and harmful type of malicious tampering. Recently, face-related security issues have attracted widespread attention due to the bad events caused by the attack of face photos/video on the ride-hailing client’s face verification and alipay’s face payment. Therefore, it is of great significance to study the forensics technology of face image authenticity.

Previous researches on image forensics focusing on digital authenticity, content authenticity and object authenticity, have made certain achievement. However, all of them have specific conditions and scope of application, and are easily attacked by anti-forensics. In practical applications, it’s very common that a variety of forensic methods are needed to give valid appraisal opinions. As an essential element of imaging, illumination has been widely studied and applied in image forensics. But there are still many places worthy of further exploration, improvement and flexible application. Therefore, the research objective of this paper is to study the face image forensics method based on illumination analysis, so as to obtain the methods with higher accuracy, better robustness and stronger persuasiveness, and apply them in practical problems.

The main research contents and innovations of this paper have the following three points: 1.The complex shadows on the human face and the error in the reconstructed face model can seriously affect the accuracy of existing illumination estimation methods. In this paper, a method of illumination estimation based on ambient light/diffuse light separation (LEAD) is proposed by using the physical fact that shadows reflect ambient light conditions objectively. Firstly, we propose a sphere harmonic lighting based on ambient/diffuse light separation (SHAD) which is more in line with physical facts. On this basis, the iterative method is employed to estimate ambient illumination and diffuse illumination. Experimental results show that compared with the existing illumination estimation methods, our proposed method enhances the robustness of illumination estimation for hair texture, complex shadows, and normal vector errors. The illumination obtained by this method, combined with the ambient light and other information for the illumination consistency judgment, offers more objective and convincing judgment opinions for image forensics. 2.The traditional method estimate the illumination based on the parallel projection and weak perspective projection, which deviates from the fact that the camera model is a perspective model. As a result, the direct comparison of consistency analysis using these methods will bring inherent errors. This paper presents a spatial illumination estimation algorithm under perspective projection. By unifying the coordinate system of each object into the camera coordinate system, the spatial illumination of each object relative to the camera coordinate system is obtained, and the accuracy of the illumination estimation method is significantly improved from the physical principle. Experimental results show that compared with the traditional illumination consistency analysis based on parallel projection illumination, the spatial illumination obtained by this method has higher accuracy. What’s more, combing with the equivalent focal length, space face location, re-projected image and other related information for photo space rationality analysis, we can get a more convincing conclusion. 3. According to the fact that the they are under the umbrella of photos and screens, the spoofing faces are unable to respond to changes in external illumination, while illumination distribution of the real faces changes with illumination, this paper proposes a face spoofing detection method based on illumination change analysis to identify real faces and spoofing faces by calculating the correlation of illumination changes of face and scene before and after external illumination changes. Through the steps of light estimation, changing illumination, calculating correlation, and hypothesis testing judgment, it is effectively perform face spoofing judgment. Experimental results show the effectiveness of proposed method on multiple kinds of face attacks including printed photo, screen photo, and video replay attack, and other advantages such as user cooperation free, loose using conditions, simple equipment demand, easy to camouflage and propitious to face recognition.

Pages144
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23937
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
张旭. 基于光照分析的人脸图像取证算法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
基于光照分析的人脸图像取证算法研究.pd(5870KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[张旭]'s Articles
Baidu academic
Similar articles in Baidu academic
[张旭]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[张旭]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.