一种改进的双流Faster R-CNN图像篡改识别模型

IMAGE FORGERY RECOGNITION MODEL BASED ON IMPROVED DUAL-STREAM FASTER R-CNN

针对目前大多数图像篡改算法只能针对一类图像篡改进行检测,以及双流Faster R-CNN算法提取的RGB流和噪声流特征对多种图像篡改检测精度不高的问题,提出一种通用的基于改进的双流Faster R-CNN图像篡改识别算法。提取图像的YCrCb颜色空间,代替之前的RGB颜色空间,以更好找出篡改的痕迹;对提取噪声特征的三个隐写分析丰富模型(SRM)滤波器进行旋转变换,以更好区别真实区域和篡改区域的噪声不一致,从而提高对篡改图像的识别精度;通过双线性池化,输入网络训练和分类,完成对图像篡改的检测与篡改区域定位。为验证算法的性能,在CASIA和NISIT16两个数据集上进行了实验。结果表明,与双流Faster R-CNN算法相比,提出的图像篡改识别算法在拼接检测、复制移动检测和移除检测上平均精度(AP)分别提升0.9百分点、1.5百分点和2.6百分点。

At present,most image tampering algorithms can only detect one type of image tampering,and the features of RGB stream and noise stream extracted by the dual-stream Faster R-CNN algorithm are not highly accurate for various image tampering detection.This paper proposes a general improved dual-stream Faster R-CNN image tampering recognition algorithm.By extracting the YCrCb color space of the image,replacing the previous RGB color space,we could find out the traces of tempering better.The three steganalysis rich model(SRM)filters for extracting noise features were rotated to better distinguish the noise inconsistency between the real area and the tampered area,so as to improve the recognition accuracy of the tampered image.Through bilinear pooling,inputting network training and classification,the detection of image tampering and the location of tampering area were completed.In order to verify the performance of the algorithm,experiments were conducted on CASIA and NISIT16 data set.The results show that,compared with the dual-stream Faster R-CNN method,the proposed algorithm improves the average precision(AP)of splicing detection,copy-move detection and removal detection by 0.9,1.5 and 2.6 percentage points respectively.