PaperlayerView exact revision on arXiv

Source-linked AI summary

Flexible-Region Based Adaptive In-Loop Filter for Video Coding

Xuewei Meng, Chuanmin Jia, Jing Cui, Shanshe Wang, Siwei Ma

arXiv:2609.18983v2eess.IVcs.CVcs.MM

TL;DR

Existing AVS3 ALF uses fixed region partitions and restricted neighboring-region merging, limiting how it accounts for texture differences and non-local similarities. FRALF introduces flexible partition templates and non-local coefficient merging; it reports average bit-rate savings across AI, RA, and LDB with marginal complexity increase.

  • Problem

    AVS3’s fixed partitioning can create uneven regions, while neighboring-region-only merging may miss non-local similarities.

    Source: A. Motivation · p. 3

  • Method

    FRALF selects among 2×4, 4×4, 4×8, and 8×8 region templates and uses optimized coefficient merging that considers neighboring and non-neighboring regions.

    Source: B. Flexible-region partition strategy · p. 3, A. Motivation · p. 3

  • Results

    0.31%, 0.36%, and 0.36% average bit-rate reductions are reported for AI, RA, and LDB, respectively, with marginal complexity increase.

    Source: §V · p. 4

  • Takeaways & Limitations

    The reported results indicate improved ALF adaptability across AI, RA, and LDB configurations with marginal complexity increase.

    Source: §V · p. 4

Abstract

from arXiv · show

Adaptive loop filter (ALF) for video coding, which is designed to minimize the mean square error between original and reconstructed samples by using Wiener-based filter, has attracted increasing attention for its significant capability in improving coding efficiency. In the second and third Audio Video Coding Standard, i.e., AVS2 and AVS3, ALF is adopted as one of the in-loop filters. In current design, each frame is divided into 16 regions at most and corresponding filter coefficients are then derived and utilized to reconstruct each region. In this paper, a flexible-region based ALF (FRALF) scheme is proposed to improve the adaptability of existing ALF in AVS3, which introduces multiple region partition templates, such as $2\times4$, $4\times4$, $4\times8$ and $8\times8$. We subsequently propose the filter coefficients merging algorithm to further improve coding efficiency by estimating the distortion level of different partition regions. The proposed FRALF can fully consider the local texture characteristics as well as non-local similarities synthetically. The experimental results show that FRALF outperforms the existing region-based ALF in AVS3 with relatively low complexity increasing.

I. INTRODUCTION

In-loop filters reduce compression artifacts; ALF uses Wiener filtering to minimize MSE and is used in AVS2 and AVS3. FRALF adapts region partitioning and coefficient merging, reporting bit-rate savings over the existing AVS3 ALF.

I. INTRODUCTION

  • ALF minimizes MSE between original and reconstructed samples using Wiener filtering, while in-loop filters reduce compression artifacts such as blocking, ringing, and blurring.Source: §I · p. 1
  • Existing AVS2 and AVS3 region-based ALF shares coefficients within regions, but fixed partitions and Hilbert-scan merging can miss within-region texture differences and similarities between non-adjacent regions.Classification-based ALF offers greater adaptability but at the expense of higher computational complexity and worse hardware friendliness.Source: §I · p. 1
  • 0.31%, 0.36%, and 0.36% BD-rate savings are reported for AI, RA, and LDB, respectively, versus original ALF in HPM-3.3.The proposed adaptive region partition scheme achieves these savings.Source: §I · p. 2
  • FRALF divides frames according to texture characteristics and merges coefficients across neighboring and non-neighboring regions to account for local texture and non-local similarity.Source: §I · p. 2

II. REVIEW OF ADAPTIVE LOOP FILTER IN AVS3

AVS3 ALF combines region partitioning, coefficient merging and Wiener-filter derivation, then applies the resulting filters to neighboring samples. This paper focuses on the first two modules.

II. REVIEW OF ADAPTIVE LOOP FILTER IN AVS3

  • AVS3 ALF comprises fixed region partitioning, Hilbert-scan coefficient merging and Wiener-filter derivation, followed by group-based filtering of neighboring samples.Source: §II · p. 2
  • ALF minimizes MSE between original and reconstructed samples using a Wiener-based filter in AVS2 and AVS3.Source: §II · p. 2
  • The paper focuses on fixed region partitioning and Hilbert-scan coefficient merging, rather than group-based filtering.Source: §II · p. 2

A. Region Partition Strategy

AVS3 partitions frames into regions aligned to CTU boundaries, but this can produce uneven region sizes. Coefficients are then derived per region and merged using a Hilbert scan with restricted merge options.

A. Region Partition Strategy

  • The existing partition can yield regions with uneven CTU counts and irregular texture distributions, potentially reducing filtering performance.For 416 × 240 video, the example has four regions; for 1280×720, it has 16, ranging from 2 to 12 CTUs.Source: A. Region Partition Strategy · p. 2

B. Hilbert-Scan Based Coefficients Merging and Coefficients Derivation

  • Pixels sharing a region index are gathered, and that region’s Wiener coefficients are calculated by minimizing MSE using neighboring samples.Source: B. Hilbert-Scan Based Coefficients… · p. 2
  • In Hilbert-scan merging, region N can merge only from region N−1, even when a non-neighboring region may have more similar characteristics.The passage gives region 1 as a case where another non-neighboring region may be more similar than region 2.Source: B. Hilbert-Scan Based Coefficients… · pp. 2–3

C. Group-based Filtering

AVS3 applies a symmetric Wiener-based filter to pixel samples using neighboring samples and the region’s filter coefficients. CTU-level control and temporal coefficient prediction address filtering and signaling overhead.

C. Group-based Filtering

  • The AVS3 filter combines a 7 × 7-tap cross with a 3 × 3-tap rectangle and filters each sample R(i, j) into R̂(i, j).K is the filter length (7), and f(k, l) denotes the filter coefficient.Source: C. Group-based Filtering · p. 3
  • A CTU-level on/off flag can prevent over-smoothing by indicating whether ALF is applied to that CTU.Source: C. Group-based Filtering · p. 3
  • Previously coded filter coefficients may be used for the current frame through temporal prediction to reduce coefficient overhead.Source: C. Group-based Filtering · p. 3

A. Motivation

The existing ALF partition template can produce imbalanced regions and may not adapt to complex high-resolution textures. Its merging scheme also restricts merges to neighboring regions, despite similarities elsewhere in the frame.

A. Motivation

  • Fixed partitioning can leave regions unbalanced and may not fit the complex textures of 4K and 8K video.The passage gives 416 × 240 and 720p examples where alternative region layouts may perform better.Source: A. Motivation · p. 3
  • Similar regions can be spatially distant, but the existing scheme restricts merging to neighbors; Region 0 is similar to regions 15, 14, and 13.The passage contrasts those similarities with Region 0's permitted merge with Region 1.Source: A. Motivation · p. 3
  • Figure 4 is an example of region partition and coefficient merging for ALF in AVS3.Source: Figure 4 · p. 3

B. Flexible-region partition strategy

The proposed partition strategy offers four region templates, selects among them by rate-distortion performance, and signals the chosen template index to the decoder.

B. Flexible-region partition strategy

C. Non-local based filter coefficients merging scheme

The merging scheme allows merges among any regions, uses fast distortion estimation to select a reliable mode, and avoids increasing the number of iterations. Template-selection statistics vary with resolution and QP.

C. Non-local based filter coefficients merging scheme

  • Merging is permitted between any regions, while fast distortion estimation identifies a reliable merging mode without increasing the iteration count.For a frame with R regions, the method performs R iterations, merging the regions into i at iteration i.Source: C. Non-local based filter coefficients… · pp. 3–4
  • High-resolution videos tend to select templates with more regions, while low-resolution videos tend to select templates with fewer regions.Source: C. Non-local based filter coefficients… · p. 4
  • At high QPs, 4×8 and 8×8 partition modes are more likely to be selected.The passage attributes this pattern to better pre-ALF image quality and more complicated texture details at high QPs.Source: C. Non-local based filter coefficients… · p. 4
  • Figure 6 reports selection probabilities for different partition templates under RA configuration.Source: Figure 6 · p. 4

IV. EXPERIMENTAL RESULTS

Under AVS3 common test conditions, FRALF reduces bitrate against HPM-3.3 with ALF enabled across AI, RA, and LDB, with minor encoding and decoding time increases. The section also reports gains against an ALF-disabled anchor and on City under LDB.

IV. EXPERIMENTAL RESULTS

  • Tests used HPM-3.3, AVS3 common test conditions, and AI, RA, and LDB configurations; coding performance was measured by Y-component BD-rate.Encoder and decoder running times were also evaluated.Source: §IV · p. 4
  • 0.31%, 0.36%, and 0.36% bitrate reduction in AI, RA, and LDB, respectively, with ALF enabled and minor encoding/decoding time increase.Source: §IV · p. 4
  • 2.87%, 4.05%, and 3.74% bitrate savings in AI, RA, and LDB, respectively, against HPM-3.3 with ALF disabled.Source: §IV · p. 4
  • On City under LDB, FRALF adds up to 1.07% coding gain over existing ALF, with total adaptive-loop-filter bitrate reduction of 7.70%.Source: §IV · p. 4
  • The reported RA and LDB bitrate reductions are higher because improved reference-picture quality benefits compression of following frames.Source: §IV · p. 4
  • Figure 7 shows rate-distortion curves for vidyo1 under LDB configuration.Source: Figure 7 · p. 4

V. CONCLUSION

The conclusion proposes a flexible-region ALF that combines multiple partition templates with non-local filter-coefficient merging. It reports bit-rate reductions across AI, RA, and LDB with marginal complexity increase.

V. CONCLUSION

  • 0.31%, 0.36%, and 0.36% average bit-rate reductions are reported for AI, RA, and LDB, respectively, with marginal complexity increase.The method is presented as improving ALF adaptability.Source: §V · p. 4
  • Multiple region-partition templates and non-local filter-coefficient merging are introduced for high-resolution and multi-scene video coding.The scheme considers local texture characteristics and non-local similarities.Source: §V · p. 4
  • The proposed flexible-region partitioning aims to improve ALF coding efficiency while considering local texture and non-local similarities.Source: §V · p. 4
Loading 2609.18983v2…