Back to Search Document details
14th Meeting: Geneva, March 2019 2019-03-19 15:29
CE13-2.6/CE13-2.7: Evaluation results of CNN based in-loop filtering
Abstract
This contribution reports the results by the convolutional neural network-based loop filtering which was proposed in JVET-M0872. Experimental results are followed by the test conditions of Core Experiment 13 (CE13) description. Simulation results reportedly show the BD-rates for luma by -1.70%, 2.26%, -0.80% and -1.68% in CE13-2.6a (CE13-2.7a), CE13-2.6b, CE13-2.6c and CE13-2.7b, respectively. The difference is caused by the CNN based filter location and CTU-level switching.
JVET-N0710 CE13-2.6/CE13-2.7: Evaluation results of CNN based in-loop filtering [K. Kawamura, Y. Kidani, S. Naito (KDDI)] [late]

The filter structure is shown in the figure blow. This filter has four layers with 3x3 taps. The input of “sum” block is residual signal from the left signal and reconstructed signal just after deblocking filter. The output of “sum” block is filtered pixels. Actual output is weighted sum of after-filtered and before-filtered pixels based on the distance of edge.

The structure of proposed CNNF

Filter coefficients are pre-defined in both the encoder and the decoder so that additional side information is not required. A representation of filter coefficients is fixed point value (integer value). The filter is applied for both intra and inter pictures.

Non-CE Technology proposals (383)

CE1 related – Post-prediction and post-reconstruction filtering (11)

Contributions in this category were discussed Thursday 21 March 1150–1400 (Track A chaired by JRO).

PATENTS:
US11562224B2 0.72 2023-01-24 US11164312B2 0.34 2021-11-02 US20230409715A1 0.32 2023-12-21 JP6928371B2 0.30 2021-09-01 CN115601772B 0.28 2023-05-02 US11837354B2 0.26 2023-12-05 US12406583B2 0.24 2025-09-02 US20200205697A1 0.20 2020-07-02
Decisions
Contributions in this category were discussed Thursday 21 March 1150–1400 (Track A chaired by JRO).
Citation