Network Working Group JM. Valin Internet-Draft Mozilla Intended status: Standards Track March 21, 2016 Expires: September 22, 2016 Directional Deringing Filter draft-valin-netvc-deringing-01 Abstract This document describes a deringing filter that takes into account the direction of edges and patterns being filtered. The filter works by identifying the direction of each block and then adaptively filtering along the identified direction. In a second pass, the blocks are also filtered in a different direction, with more conservative thresholds to avoid blurring edges. The proposed deringing filter is shown to improve the quality of both Daala and the Alliance for Open Media (AOM) video codec. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at http://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on September 22, 2016. Copyright Notice Copyright (c) 2016 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must Valin Expires September 22, 2016 [Page 1] Internet-Draft Deringing March 2016 include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License. This document may not be modified, and derivative works of it may not be created, and it may not be published except as an Internet-Draft. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 2. Direction Search . . . . . . . . . . . . . . . . . . . . . . 2 3. Directional Filter . . . . . . . . . . . . . . . . . . . . . 3 4. Second Stage Filter . . . . . . . . . . . . . . . . . . . . . 4 5. Setting Thresholds . . . . . . . . . . . . . . . . . . . . . 5 6. Superblock Filtering . . . . . . . . . . . . . . . . . . . . 6 7. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 8. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 7 9. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 7 10. Security Considerations . . . . . . . . . . . . . . . . . . . 7 11. Informative References . . . . . . . . . . . . . . . . . . . 7 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 8 1. Introduction This document describes a deringing filter that takes into account the direction of edges and patterns being filtered. The filter works by identifying the direction of each block and then adaptively filtering along the identified direction. In a second pass, the blocks are also filtered in a different direction, with more conservative thresholds to avoid blurring edges. The deringing filter is implemented for both Daala and the Alliance for Open Media (AOM) codec. 2. Direction Search The first step is to divide the image into blocks of fixed or variable size. Variable-size blocks make it possible to use large blocks on long, continuous edges and small blocks where edges intersect or change direction. A fixed block size is easier to implement and does not require signaling the sizes on a block-by- block basis. For this work, we consider a fixed block size of 8x8. Once the image is divided into blocks, we determine which direction best matches the pattern in each block. One way to determine the direction is to minimize mean squared difference (MSD) between the input block and a perfectly directional block. A perfectly directional block is a block for which each line along a certain Valin Expires September 22, 2016 [Page 2] Internet-Draft Deringing March 2016 direction has a constant value. For each direction, we assign a line number to each pixel, as shown below. +---+---+---+---+---+---+---+---+ | 0 | 0 | 1 | 1 | 2 | 2 | 3 | 3 | +---+---+---+---+---+---+---+---+ | 1 | 1 | 2 | 2 | 3 | 3 | 4 | 4 | +---+---+---+---+---+---+---+---+ | 2 | 2 | 3 | 3 | 4 | 4 | 5 | 5 | +---+---+---+---+---+---+---+---+ | 3 | 3 | 4 | 4 | 5 | 5 | 6 | 6 | +---+---+---+---+---+---+---+---+ | 4 | 4 | 5 | 5 | 6 | 6 | 7 | 7 | +---+---+---+---+---+---+---+---+ | 5 | 5 | 6 | 6 | 7 | 7 | 8 | 8 | +---+---+---+---+---+---+---+---+ | 6 | 6 | 7 | 7 | 8 | 8 | 9 | 9 | +---+---+---+---+---+---+---+---+ | 7 | 7 | 8 | 8 | 9 | 9 |10 |10 | +---+---+---+---+---+---+---+---+ For each direction d, we compute the value s_d, which is equal to a direction-independent offset minus the MSD (see [Deringing-Note] for detauls) as: __ __ 2 \ 1 / \ \ s_d= /_ ------- * | /_ x_p | , k in block N_(d,k) \ p in P_(d,k) / where x_p is the value of pixel p, P_(d,k) is the set of pixels in like k along direction d, and N_(d,k) is the cardinality of P_(d,k). From there, the direction is computed as the value of d that maximizes s_d. 3. Directional Filter The directional filter for pixel (i,j) is defined as the 7-tap non- linear filter Valin Expires September 22, 2016 [Page 3] Internet-Draft Deringing March 2016 3 _ 1 -- | / \ y(i,j)=x(i,j)+---*\ w_k*| f| x(i,j)-x(i+floor(k*d_y),j+floor(k*d_x), T| W /_ |_ \ / k=1 _ / \ | + f| x(i,j)-x(i-floor(k*d_y),j-floor(k*d_x), T| | \ /_| where d_x and d_y define the direction, W is a constant normalizing factor, T is the filtering threshold for the block, and f(d,T) is defined as / \ d , |d| < T f(d, T) = < / 0 , otherwise \ The direction parameters are shown in the table below. The weights w_k can be chosen so that W is a power of two. For example, Daala currently uses w=[3 2 2] with W=16. Since the direction is constant over 8x8 blocks, all operations in this filter are directly vectorizable over the blocks. +-----------+------+------+ | Direction | d_x | d_y | +-----------+------+------+ | 0 | 1 | -1 | | 1 | 1 | -1/2 | | 2 | 1 | 0 | | 3 | 1 | 1/2 | | 4 | 1 | 1 | | 5 | 1/2 | 1 | | 6 | 0 | 1 | | 7 | -1/2 | 1 | +-----------+------+------+ Table 1 4. Second Stage Filter The 7-tap directional filter is sometimes not enough to eliminate all ringing, so we use an additional filtering step that operates across the direction lines used in the first filter. Considering that the input of the second filter has considerably less ringing than the input of the second filter, and the fact that the second filter risks Valin Expires September 22, 2016 [Page 4] Internet-Draft Deringing March 2016 blurring edges, the position-dependent threshold T_2(i,j) for the second filter is set lower than that of the first filter T. The filter structure is the same as the one used for the directional filter. The direction parameters for the second stage filter are shown in the table below and the filter weights are w=[1 1] with W=16/3. +-----------+-----+-----+ | Direction | d_x | d_y | +-----------+-----+-----+ | 0 | 1 | 0 | | 1 | 0 | 1 | | 2 | 0 | 1 | | 3 | 0 | 1 | | 4 | 1 | 0 | | 5 | 1 | 0 | | 6 | 1 | 0 | | 7 | 1 | 0 | +-----------+-----+-----+ Table 2 5. Setting Thresholds The thresholds T and T_2 must be set high enough to smooth out ringing artefacts, but low enough to avoid blurring important details in the image. Although the ringing is roughly proportional to the quantization step size Q, as the quantizer increases the error grows slightly less than linearly because the unquantized coefficients become very small compared to Q. As a starting point for determining the thresholds, Daala uses a power model of the form T_0=level*alpha_1*Q^beta, with beta=0.842, and where alpha_1 depends on the input scaling. The level is a threshold adjustment coded for each superblock (64x64). In the AOM codec, a global threshold is selected by the encoder instead of using a function of the quantizer, so T_0=level*global_level. Another factor that affects the optimal filtering threshold is the presence of strong directional edges/patterns. These can be estimated from the s_d parameters computed in the direction search as delta=s_(d_opt)-s_(d_ortho), where d_ortho=d_opt+4 (mod 8). We compute the direction filtering threshold for each block as / 1 / 1/6 \ \ T = T_0*max| ---, min| 3, alpha_2*(delta) | |, \ 2 \ / / Valin Expires September 22, 2016 [Page 5] Internet-Draft Deringing March 2016 where alpha_2 also depends on the input scaling. For the second filter, we use a more conservative threshold that depends on the amount of change caused by the directional filter. / T \ T_2(i,j) = min| T, --- + |y(i,j)-x(i,j)| |. \ 3 / As a special case, when the pixels corresponding to the 8x8 block being filtered are all skipped, then T=T_2=0, so no deringing is performed. 6. Superblock Filtering The filtering is applied one superblock at a time, in a way that depends on the level. In Daala, the level can take one of 6 values: 0, 0.5, 0.7, 1.0, 1.4, 2.0, where a level of zero disables the deringing filter for the current superblock. The level is the only information coded in the bitstream by the deringing filter. On keyframes, it is entropy-coded based on the neighbor values. On inter-predicted frames, the level is only coded for superblocks that are not skipped and is entropy-coded based on a single adapted probability distribution (no context from the neighbors). Superblocks where no level is coded have deringing disabled. Similarly, any skipped block within a superblock has deringing disabled, even if it is signaled enabled for the superblock. The level of the deringing filter in AOM is handled similarly, except that only four levels are currently available and there is no entropy coding yet. The deringing process sometimes reads pixels that lie outside of the superblock being processed. When these pixels belong to another superblock, the filtering always uses the unfiltered pixel values -- even for the second stage filter -- so that no dependency is added between the superblocks. This makes it possible -- in theory -- to filter all superblocks in parallel. When the pixels used for a filter lie outside of the viewable image, we set f(d,T)=0. 7. Results The deringing filter described here has been implemented for the Daala [Daala-website] codec. It is available from the Daala Git repository [Daala-Git]. We tested the deringing filter on the Are We Compressed Yet [AWCY] ntt-short1 set over the 0.025 bit/pixel to 0.1 bit/pixel range, corresponding to a 1080p30 bitrate of 1.5 Mbit/s to 6 Mbit/s. The Bjontegaard-delta [I-D.daede-netvc-testing] rate reduction over that range was 6.5% for PSNR, 4.7% for PSNR-HVS, 5.6% Valin Expires September 22, 2016 [Page 6] Internet-Draft Deringing March 2016 for SSIM and -6.0% (regression) for FAST-SSIM. Visual inspection confirmed that the quality is indeed improved, despite the regression in the FAST-SSIM result. In AOM for the ntt-short1 set, the medium bitrate (0.02 to 0.06 bit/ pixel) Bjontegaard-delta improvement is 2.5% for PSNR, 1.5% for PSNR- HVS, 1.5% for SSIM, and -3.8% (regression) on FAST-SSIM. The high bitrate (0.06 to 0.2 bit/pixel) Bjontegaard-delta improvement is 2.0% for PSNR, 0.8% for PSNR-HVS, 1.3% for SSIM, and -3.1% (regression) on FAST-SSIM. The smaller improvement for AOM compared to Daala may be due to the newly integrated code not being mature, but also to the fact that some features in Daala tend to cause more ringing. These features include lapped transforms, quantization matrices, perceptual vector quantization, overlapped block motion compensation (OBMC), and activity masking. 8. Conclusion We have demonstrated an effective algorithm to remove ringing artefacts from coded images and videos. The proposed filter takes into account the directionality of the patterns it is filtering to reduce the risk of blurring. 9. IANA Considerations This document makes no request of IANA. 10. Security Considerations This draft has no security considerations. 11. Informative References [AWCY] "Are We Compressed Yet?", Xiph.Org Foundation , . [Daala-Git] "Daala Git repository", Xiph.Org Foundation , . [Daala-website] "Daala website", Xiph.Org Foundation , . Valin Expires September 22, 2016 [Page 7] Internet-Draft Deringing March 2016 [Deringing-Note] Valin, JM., "The Daala Directional Deringing Filter", arXiv:1602.05975 [cs.MM] , . [I-D.daede-netvc-testing] Daede, T. and J. Jack, "Video Codec Testing and Quality Measurement", draft-daede-netvc-testing-02 (work in progress), October 2015. Author's Address Jean-Marc Valin Mozilla 331 E. Evelyn Avenue Mountain View, CA 94041 USA Email: jmvalin@jmvalin.ca Valin Expires September 22, 2016 [Page 8]