Multi-task and multi-rate feature encoding and decoding methods, systems, devices and media

CN123123592APending Publication Date: 2026-11-17UNIV OF SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611303436.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing multi-task, multi-rate feature coding schemes require training feature codecs separately for different tasks and different bitrates, resulting in high training and storage overhead. Furthermore, independent codebooks are difficult to share across different tasks and bitrates, and there is a lack of effective methods to evaluate the global adaptability of candidate feature sets.

Method used

The source feature set is selected by global fitness evaluation, and rate-distortion loss is used to jointly optimize the shared vector quantization codebook and discrete entropy model. A lightweight adapter is used to handle the differences between different tasks and bitrates, so as to realize the sharing of the same vector quantization codebook and entropy model among multiple tasks and multiple bitrates.

Benefits of technology

It reduces the training, storage, and deployment overhead of multi-task, multi-bitrate edge-cloud collaborative systems, improves the system's scalability and maintenance efficiency, and does not change the original feature extraction network and downstream task network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123592A_ABST
    Figure CN123123592A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-task and multi-rate feature encoding and decoding method, system, device, and medium, which are corresponding solutions. In these solutions: the training source of the shared codebook is selected through global feature adaptability, improving the adaptability of the shared codebook to intermediate features of different tasks and bitrates; multiple tasks and multiple bitrates share the same vector quantization codebook and discrete entropy model, reducing the parameters generated by configuring independent codebooks and entropy models for different task-bitrate combinations, thus lowering the overhead of model training, storage, and deployment; overall, the above solutions unify global adaptability evaluation, shared vector quantization codebook, and lightweight task-bitrate adaptation mechanism into the encoding process of multi-task and multi-bitrate intermediate features, enabling intermediate features of different tasks to complete quantization, entropy encoding, entropy decoding, and feature reconstruction using the same shared codebook, thereby reducing the training, storage, deployment, and expansion overhead of the edge-cloud collaborative system.
Need to check novelty before this filing date? Find Prior Art