Back to Search Document details
22nd Meeting: by teleconference, April 2021 2021-04-22 17:09
EE1-1.1: neural network based in-loop filter using depthwise separable convolution and regular convolution (JVET-U0061)
Authors: Zishan Li Tencent Cheung Auyeung Tencent Xiaozhong Xu Tencent W. Wang Tencent Xinwei Li Tencent Shan Liu Tencent
Abstract
This contribution presents the updated result of the neural-network-based in loop filter using depthwise separable convolution (DSC) in EE1.1 with new test condition that GOP=32, and QP= {22,27,32,37,42}. With VTM-11.0 as anchor, on average results from RA/AI configurations, report gains for luma BD-Rate are 1.15% and 1.61% respectively.
Meeting Notes
JVET-V0137 EE1-1.1: neural network based in-loop filter using depthwise separable convolution and regular convolution (JVET-U0061) [Z. Li, C. Auyeung, X. Xu, W. Wang, X. Li, S. Liu (Tencent)] [late]

This contribution presents the updated result of the neural-network-based in loop filter using depthwise separable convolution (DSC) in EE1.1 with new test condition that GOP=32, and QP= {22,27,32,37,42}. With VTM-11.0 as anchor, on average results from RA/AI configurations, report gains for luma BD-Rate are 1.15% and 1.61% respectively.

Filter is a new stage between deblocking and SAO.

QP information is input to the network.

Only one model was used, with no local (e.g. CTU based) control.

References:
JVET-T0057 JVET-U0060 JVET-U0061
Cited:
JVET-V0011 JVET-V0023 JVET-V2023 JVET-W0023 JVET-W0151
PATENTS:
20220191553 0.34 WO/2022/132277 0.33 20220210446 0.32 WO/2022/146503 0.31 WO/2024/010710 0.30 20220405545 0.28 WO/2022/261968 0.27 20210004589 0.26
ARCHIVE v3 :
.DS_Store JVET-V0137-v1.pptx JVET-V0137-v2.docx JVET-V0137-v2.xlsm
Decisions
Only one model was used, with no local (e.g. CTU based) control.
Citation