Back to Search Document details
37th Meeting: Geneva, CH, January 2025 2025-01-16 18:21
[AHG11] A Hybrid Framework Integrating End-to-End Learned Image Codec with Conventional Codec
Authors: Nannan Zou Nokia Antti Hallapuro Nokia Francesco Cricri Nokia Honglei Zhang Nokia Miska M. Hannuksela Nokia
Abstract
This contribution proposes a hybrid framework that integrates end-to-end learned image compression (LIC) methods with conventional compression techniques. The framework involves using LIC-coded intra frames and VTM-coded inter frames. Furthermore, the encoder decides whether to use the LIC-coded intra frames or the VTM-coded intra frames. With this hybrid framework, under the Random-Access configuration, the resulting BD-rates over NNVC-7.1 VTM (with NN tools off) anchor are reported to be as follows:
Meeting Notes
JVET-AK0146 [AHG11] A Hybrid Framework Integrating End-to-End Learned Image Codec with Conventional Codec [N. Zou, A. Hallapuro, F. Cricri, H. Zhang, M. M. Hannuksela (Nokia)]

This contribution proposes a hybrid framework that integrates end-to-end learned image compression (LIC) methods with conventional compression techniques. The framework involves using LIC-coded intra frames and VTM-coded inter frames. Furthermore, the encoder decides whether to use the LIC-coded intra frames or the VTM-coded intra frames. With this hybrid framework, under the Random-Access configuration, the resulting BD-rates over NNVC-7.1 VTM (with NN tools off) anchor were reported to be as follows:

Class A1 0,00% (Y), 0,00% (Cb), 0,00% (Cr)

Class A2 -0,27 % (Y), -0,34 % (Cb), -0,35 % (Cr)

Class B -0,18 % (Y), -8,81 % (Cb), -8,35 % (Cr)

Class C 0,22 % (Y), -7,60 % (Cb), -7,32 % (Cr)

Class D -0,70 % (Y), -11,48 % (Cb), -11,34 % (Cr)

This uses a very large network, with 3.3 MMAC/s encoder, 8.5 for decoder; the exact architecture was not described.

Training uses BVI, DIV2K, JPEG-AI, CLIC and some other

Proponents are requested to upload the presentation (in v2 another version was provided)

Encoder can switch between conventional and learned method for the entire I frame.

Further study was highly encouraged. A more precise description should be provided, and complexity reduction, with a reduction of number of models, was recommended.

ARCHIVE :
AI.xls.xlsm JVET-AK0146.docx JVET-AK0146_v2.pptx RA.xls.xlsm
Decisions
Further study was highly encouraged. A more precise description should be provided, and complexity reduction, with a reduction of number of models, was recommended.
Citation