Combining Progressive Rethinking and Collaborative Learning: A Deep Framework for In-Loop Filtering

In this paper, we aim to address issues of (1) joint spatial-temporal modeling and (2) side information injection for deep-learning based in-loop filter. For (1), design a deep network with both progressive rethinking collaborative learning mechanisms improve quality the reconstructed intra-frames inter-frames, respectively. intra coding, Progressive Rethinking Network (PRN) is designed simulate human decision mechanism effective spatial modeling. Our block introduces an additional inter-block connection bypass high-dimensional informative feature before bottleneck module across blocks review complete past memorized experiences rethinks progressively. inter current frame interacts reference frames (peak nearest adjacent frame) collaboratively at level. (2), extract intra-frame inter-frame better context A coarse-to-fine partition map on HEVC trees built as information. Furthermore, warped features are offered PRN provides 9.0% BD-rate reduction average compared baseline under All-intra (AI) configuration. While Low-Delay B (LDB), P (LDP) Random Access (RA) configuration, our 9.0%, 10.6% 8.0% project webpage https://dezhao-wang.github.io/PRN-v2/.