基于K-SVD字典学习的合成图像盲检测

Blind Detection of Composite Images Based on K-SVD Dictionary Learning

一幅真实的图像中噪声特性是一致的,而由多幅图像内容拼接而成的合成图像噪声特性是不一致的.本文利用这一特点,提出了一种基于K均值奇异值分解(K-SVD)字典学习的合成图像盲检测方法.该方法首先通过K-SVD算法对合成图像进行训练得到其稀疏表示字典,然后利用学习得到的字典对背景噪声进行去除,最后根据去噪前后图像对应子块的相关系数异同实现篡改区域的检测与定位.实验结果表明,该方法对于鉴别含有不同背景噪声的合成图像具有显著效果,同时,算法对JPEG压缩、重采样和模糊等后处理操作都具有较好的鲁棒性.

Digital images which have an inherent amount of noise typically uniform across the entire image is introduced by imaging process,if images with different noise levels are spliced together would leave an evidence of tampering.Base on this characteristic,in this paper we proposes a blind detection method using K-means singular valne deconposition（K-SVD）dictionary learning.A dictionary about sparse representation is obtained by training samples form composite image with K-SVD dictionary learning algorithm.Then the composite image is denoised by utilizing learned dictionary.By estimating the correlation coefficients of image blocks before and after denoise,the fogery regions can be found.Simulation results show its effectiveness in detecting forgery part in spliced images with different noise levels.The proposed method has good robustness against lossy JPEG compression,resampling and blurring.