A local feature based compensation scheme for unsmooth video coding

In this paper, a local feature based compensation scheme is proposed for unsmooth video coding. Unlike natural video sequences, video sequences captured by mobile device, often have unsmooth changes, which makes the scene content have a sharp change in geometric deformation and photometric transformation. Unfortunately, the unsmooth changes are often difficult to be reduced by block motion compensation. Different from the method used in the traditional video coding scheme, this paper proposes a novel local feature based compensation scheme utilizing the technology of multi-reference frames motion compensation. Firstly the geometric distortions are eliminated through geometric model estimation utilizing the correlations between the current encoding frame and the frames in reference buffers. Then we obtain the deformed prediction frame through the photometric transformation module to compensate the illumination changes. Finally, the predicted frames are substituted into the reference buffers for video encoding. Experiments show that the proposed method can achieve better video quality with lower stream compared with the traditional method.