基于透视投影下空间光照一致性分析的图像拼接篡改检测

将一个人的头像剪切并拼接到另一张照片中，是一种常见的图像篡改手段.如果将该合成照片用于敲诈勒索，会对社会带来严重危害.因此，用来检测图像篡改的图像取证技术具有重大意义.由于不同照片成像环境不同，拼接时很难做到不同人脸的光照绝对一致，因此可以通过光照是否一致检测篡改.以往光照估计方法基于平行投影的假设，利用照片投影光照进行光照一致性分析.实际上，相机针孔模型是透视投影，从而导致上述检测方法出现误差.针对这一问题，本文提出一种透视投影下物体空间光照估计算法，将各人脸姿态统一到相机坐标系下，估计各人脸相对于相机坐标系的空间光照，然后分析空间光照一致性.另外，根据人脸空间光照一致性约束可以优化出相机参数，并得到该参数下的等效焦距、人脸空间位置及重新透视投影的图像等空间信息.本文将空间光照的一致性和上述空间信息的合理性作为依据，对人脸图像进行拼接篡改检测.实验结果表明，相比于传统方法基于平行投影光照进行光照一致性分析，采用本文提出的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，结合相关信息进行照片空间合理性分析的篡改检测方法具有更强的说服力.

Image Splicing Detection Based on Spatial Lighting Consistency Analysis Under Perspective Projection

Splicing a face image of a person into another photo is a common tempering way, and it is harmful once used for blackmail. Therefore, the image forensics for tampered image detection is of great significance. As different photos are taken in different environments with different cameras, it is very difficult to ensure the spliced faces have absolutely consistent light. Thus the illumination consistency can be an effective clue for tempering detection. Existing illumination consistency analysis based tempering detection methods perform the illumination estimation under the assumption of parallel projection. However, the results usually have huge but subtle errors because the camera pinhole model is perspective projection. To solve this problem, we propose a spatial lighting estimation method under perspective projection for the illumination consistency analysis. In addition, the spatial information analysis is combined to obtain more reasonable detection results. Specifically, the coordinate system of each face is firstly unified into the camera coordinate system and the spatial lighting of each face relative to the camera coordinate system is estimated. Then the camera parameters are optimized according to the consistency constraint of faces' spatial lighting and the spatial information including the equivalent focal length, face spatial position and re-perspective projection image under the camera parameters are calculated. Finally, the consistency of spatial lighting and the rationality of the spatial information are both taken as clues to detect face image splicing forgery. The experimental results show that compared with previous methods, the spatial lighting estimation method has higher accuracy, and the spatial information analysis method has stronger persuasion.