Content adaptive update for lifting-based motion-compensated temporal filtering - Electronics Letters

Introduction: Lifting-based motion-compensated temporal filtering (MCTF) is used widely as the preferred temporal transformation technique in state-of-the-art three-dimensional subband video coding. It involves two basic stages: the prediction stage that takes the original input frames to generate the highpass frames and the update stage that uses the available highpass frames and even frames to generate the lowpass frames. Each highpass frame is essentially the residual from motion-compensated prediction of the relevant odd indexed original video frames. Its energy depends only upon the success of the motion models. When the motion model fails, object features may not be aligned, resulting in multiple edges and increased energy in the highpass subband frames, which not only reduce the compression performance, but also add ghosting to lowpass temporal frames, significantly reducing their visual quality. Turaga [1] proposed omitting the update steps based on quality of motion estimation and the nature of the motion in the sequence. Mehrseresht [2] showed that ignoring the update not only incurs the loss of compression performance, but also increases the fluctuation in PSNR values. Therefore, we should mitigate adaptively the ghosting artefacts of lowpass temporal frames, while preserving the update steps in general.