Estimating QP and motion vectors in H.264/AVC video from decoded pixels

Abstract

In this paper we present a method for blindly estimating the quantization parameter (QP) and the motion-vectors in H.264/AVC decoded video. We assume that only the decoded pixel values are available. This models pretty well the dissemination over the Internet of user-generated contents, where a video may pass through several coding/processing stages from acquisition to publication and download. The proposed technique may be a relevant tool in the forensics field, as it can help to reconstruct the previous history of the digital content, or to provide evidence of tampering. Most of the previous work in this direction simply adapts still image forensic approaches to video, but their applicability is in general limited (e.g.\ they can deal with Intra frames only). Conversely, we explicitly take into consideration motion-compensated prediction used by state-of-the-art video codecs such as H.264/AVC to find the QP also for P frames. We show that the so-obtained QPs can be used to estimate the original motion field of the encoder.