Synthetic Transform Network With Cascaded Attention for Video Decoding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural network-based encoding and decoding methods face challenges with poor decoding performance and high complexity in video coding processes.

Innovation Solution

Implement a synthetic transform network with a cascaded hybrid attention module to reduce network and calculation complexity while improving decoding performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If neural network-based encoding and decoding methods are used, then decoding performance can be improved, but system complexity increases significantly

Engineering Contradiction:
Improvedecoding performanceVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent divides the complex neural network into separate encoding and decoding modules, each handling specific tasks. The encoder processes video data through multiple layers to generate compressed representations, while the decoder reconstructs the video from these representations. This segmentation allows each module to be optimized independently, improving overall decoding performance while managing system complexity through modular design.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces intermediate representations (feature maps, latent codes) as mediators between the encoder and decoder. These intermediate structures serve as bridges that transform complex video data into compact forms suitable for transmission and reconstruction, reducing the direct complexity burden on both encoding and decoding operations while maintaining high decoding performance.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Manufacturing precision

If complex neural network structures are employed to improve decoding quality, then reconstruction accuracy increases, but computational complexity increases

Engineering Contradiction:
Improvereconstruction accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent employs dynamic computational structures where the network adapts its processing based on input characteristics. The encoder and decoder use learnable parameters that are optimized during training to handle varying video content, allowing high reconstruction accuracy to be achieved dynamically rather than through static complex structures, thus managing computational complexity more efficiently.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent utilizes parameter optimization and adjustment as key mechanisms. During training, network parameters are continuously adjusted to maximize reconstruction accuracy. The system changes parameters adaptively based on the statistical properties of the video data, enabling high reconstruction quality without requiring permanently complex computational structures, thereby reducing overall computational complexity.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4708860A1Decoding and encoding method, apparatus, and device thereof
Publication Date: 2026.03.11 HANGZHOU HIKVISION DIGITAL TECHNOLOGY CO LTD
  • EP4708860A1 patent drawingFigure 1~2
  • EP4708860A1 patent drawingFigure 3~4
  • EP4708860A1 patent drawingFigure 5~6A

AI summary

The present application provides a decoding and encoding method, an apparatus, and a device thereof. The decoding method comprises: decoding a code stream corresponding to a current image block to obtain coefficient hyperparametric features; on the basis of the coefficient hyperparametric features, determining probability distribution parameters, and on the basis of the probability distribution parameters, decoding another stream of the current image block to obtain reconstructed features; inputting the reconstructed features into a synthetic transformation network to obtain reconstructed image blocks; wherein the synthetic transformation network comprises an attention module, and the attention module is a cascaded mixed attention module; the cascaded mixed attention module comprises a first attention submodule and a second attention submodule, and the first attention submodule and the second attention submodule are connected in series.