Back to Search Document details
34th Meeting: Rennes, FR, April 2024 2024-04-18 20:23
EE1-related: Complexity reduction of NN in-loop filters through early cropping
Authors: J. Ström Ericsson Mitra Damghanian Ericsson Du Liu Ericsson P. Wennersten Ericsson
Abstract
This contribution proposes to gradually reduce the patch size in the last convolutional layers in the neural network-based in-loop filters to avoid unnecessary computation. This way, the output patch from the network can have the desired size, and subsequent cropping is not needed. It is alleged that the proposed method can reuse existing models without the need of retraining. Using the EE1-1.3 as an example, the complexity can reportedly be reduced by -0.8% to 16.76 kMAC/pixel, compared to the 16.90 kMAC/pixel of EE1-1.3. The results are claimed to be bit-exact identical to those of EE1-1.3 (no BDR impact). The reported results over EE1-1.3 are hence
Cited:
JVET-AH0334 X
ARCHIVE :
._JVET-AH0195_complexity.xlsx ._JVET-AH0195_v2_EE1-1.3-based_version_over_nnvc80.xlsm ._JVET-AH0195_v3_LOP2-based_version_over_nnvc80.xlsm ._JVET-AH0195_v4.docx ._JVET-AH0195_v4.pptx ._JVET-AH0195_v4_clean.docx ._JVET-AH0195_v4_complexity.xlsx ._JVET-AH0195_v4_EE1-1.3-based_version_over_ee113.xlsm JVET-AH0195_complexity.xlsx JVET-AH0195_v2_EE1-1.3-based_version_over_nnvc80.xlsm JVET-AH0195_v3_LOP2-based_version_over_nnvc80.xlsm JVET-AH0195_v4.docx JVET-AH0195_v4.pptx JVET-AH0195_v4_clean.docx JVET-AH0195_v4_complexity.xlsx JVET-AH0195_v4_EE1-1.3-based_version_over_ee113.xlsm
Citation