基于图像噪声残差的数字图像来源取证

针对使用支持向量机的图像来源取证算法中存在的,例如所需训练集较大(数千幅)、特征维数较高等问题。文中提出了一种取证算法,该算法仅需要少量(约为10幅)训练图像,且只提取图像噪声的残差作为图像的唯一特征。该算法首先使用小波滤波器提取图像噪声,然后借助回归模型提取噪声的残差,最后为噪声的残差建立高斯分布模型,根据不同类型图像噪声的残差模型参数进行来源取证。实验结果表明,在误报率为1.2% 的条件下,该取证算法对自然图像的准确率为95.33%,对计算机生成图像的准确率为96.44%。

Digital Image Source Forensics Based on Image Noise Residual

There are some problems with the image source forensics algorithm using the support vector machine, such as the large training set (about thousands) and the high dimension feature. To solve these problems, a forensic algorithm was proposed, which required only a small number (about 10) of training images, and only extracted the residual of the image noise as the unique feature of the image. The algorithm firstly used the wavelet filter to extract the image noise, and then extracted the residual of the image noise using the regression model. Finally, the Gaussian distribution model was established for the noise residual, and the source evidence was obtained according to different types of images with different model parameters. The experimental results showed that under the condition of FPR of 1.2%, the TPR of the proposed algorithm for natural images was 95.33%, and the TPR for computer-generated graphics was 96.44%.