Methods, systems, and training methods for encoding and decoding images, videos, and videos.

A neural network-based method for encoding and decoding images and videos using encoder-trained and hyperencoder-trained neural networks with implicit coding solvers addresses inefficiencies in existing compression methods, reducing computational cost and time while optimizing network demands.

JP7876519B2Active Publication Date: 2026-06-19INTERDIGITAL VC HOLDINGS INC
View PDF 4 Cites 0 Cited by

Patent Information

Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
INTERDIGITAL VC HOLDINGS INC
Filing Date
2021-10-25
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

The increasing demand for high-resolution and low-distortion image and video content is straining communication networks, leading to higher energy consumption and costs, while existing compression methods are inefficient and time-consuming.

Method used

A neural network-based method for encoding and decoding images and videos using encoder-trained and hyperencoder-trained neural networks, combined with implicit coding solvers, to generate and process latent representations efficiently, reducing computational cost and time.

Benefits of technology

The method significantly reduces encoding and decoding times, lowers energy consumption, and optimizes network demands by minimizing the computational load on AI-based compression pipelines.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007876519000130
    Figure 0007876519000130
  • Figure 0007876519000131
    Figure 0007876519000131
  • Figure 0007876519000132
    Figure 0007876519000132
Patent Text Reader

Abstract

1. A computer-implemented method for lossy or lossless image or video compression and transmission, the method comprising: (i) receiving an input image; (ii) encoding the input image using an encoder-trained neural network to generate a y latent representation; (iii) encoding the y latent representation using a hyper-encoder-trained neural network to generate a z hyper-latent representation; (iv) quantizing the z hyper-latent representation using a predetermined entropy parameter to generate a quantized z hyper-latent representation; (v) entropy encoding the quantized z hyper-latent representation into a first bitstream using the predetermined entropy parameter; and (vi) processing the quantized z hyper-latent representation using a hyper-decoder-trained neural network to generate a position entropy parameter μ of the y latent representation. y , the entropy scale parameter σ y , and the context matrix A y (vii) using an implicit coding solver to obtain a latent representation of y, a location entropy parameter μ y , and the context matrix A y to obtain a quantized latent residual; and (viii) an entropy scale parameter σ y and (ix) entropy encoding the quantized latent residual into a second bitstream using the quantized latent residual as a first bitstream and a second bitstream. Related computer-implemented methods, systems, computer-implemented training methods and computer program products are disclosed.
Need to check novelty before this filing date? Find Prior Art