ScholarWorks@숙명여자대학교

초록

In July 2020, Versatile Video Coding (VVC/H.266) has been finalized as the next-generation video coding standard. Due to the diverse characteristics of video, motion prediction in fractional precision is required in the video coding. For this, VVC/H.266 uses Discrete Cosine Transform-based Interpolation Filter (DCTIF) but, it is being a typical low-pass filter with fixed integer coefficients so it cannot guarantee optimal performance across all videos. Recently, deep learning-based technology has been continually developed onwards. This paper proposed the In-Loop Interpolation Filter (ILIF) which can generate high-quality fractional pixels. ILIF is an Super-Resolution (SR) model with our proposed pixel embedding technique. Pixel Embedding allows the correlation between integer and sub-pixels to be maintained during learning and it is highly effective in the inter coding. Optimized through a divide-and-conquer learning approach, ILIF replaces the DCTIF and is integrated with inter prediction in VVC/H.266. ILIF considered only the Y component of YUV420 format and the BD-rate performance was compared and analyzed with the anchor of VVC/H.266. Two integration methods (MODE 1, 2) between ILIF and VVC/H.266 were presented. As a result of the experiment, for MODE 1 which applies ILIF only for fractional pixel generation, the gains were −1.42% for All-QP, −1.54% for High-QP, and −1.24% for Low-QP. Additionally, in MODE 2 which integrates integer pixel filtering and sub-pixel generation with ILIF, it showed the gains of −3.92% for All-QP, −4.01% for High-QP, and −3.13% for Low-QP.

키워드