A confidence map and pixel-based weighted correlation for PRNU-based camera identification

In this paper, we propose a confidence map and a pixel-based weighted correlation method for digital camera identification. In traditional camera identification method, a simple denoising technique is used to extract the photo-response non-uniformity (PRNU) as the difference between the original image and the denoised image. One of the major problems is that the image content is left behind in the noise residue which affects the correlation calculation for identifying the source camera. In order to solve the problem, we first study the image content effect by examining the effect of different image features on correlation. We then formulate the image content effect by nonlinear regression model and apply this relationship to obtain a confidence map for the testing images. In such way, the confidence map shows the reliability of each pixel in correlation calculation. It can be used as a weighting function so as to give a higher weighting to pixel which is more reliable and a lower weighting for less reliable pixel. By using the weighted correlation, our proposed method is able to improve the identification accuracies especially for the low quality image such as heavy JPEG compression. In particular, we found 5–20% improvement in identification accuracy at JPEG quality factor of 70.