Video-Specific Dictionary Learning for Implicit Neural Compression

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video compression techniques face challenges in achieving high efficiency and low computational complexity, particularly in the context of two-dimensional and three-dimensional scenes or objects, as they often rely on block-based prediction and transform methods that do not leverage the full potential of neural networks.

Innovation Solution

The proposed method employs implicit neural representations (INR) with a video-specific dictionary to decompose neural network parameters into head and tail layers, approximating the head layer using a video-specific dictionary and encoding sparse coefficients and the tail layer into a bitstream, allowing for efficient compression and reconstruction.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If block-based prediction and transform methods are used for video compression, then the compression can be processed efficiently, but the compression efficiency is limited and computational complexity remains high

Engineering Contradiction:
Improvecompression efficiencyVSAvoidcomputational complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent replaces traditional block-based mechanical prediction methods with implicit neural representations that use continuous function approximations. This substitution enables the system to model video content more efficiently, achieving higher compression ratios while reducing the computational burden of repeated block processing operations

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent transforms the representation parameters from discrete block-based metrics to continuous implicit neural network parameters. By encoding and decoding these neural parameters, the system achieves better compression efficiency while the parameter optimization reduces overall computational complexity compared to traditional methods

Inventive Principle:
Principle #35Parameter changes

2Productivity

If traditional video compression methods are used, then the processing is straightforward, but they do not leverage the full potential of neural networks for better compression

Engineering Contradiction:
Improvecompression efficiencyVSAvoidimplementation complexity
Core Design Contradiction:
ProductivityVSEase of manufacture

Solution Approach 1:

The patent segments the neural network parameters into head and tail layers, allowing selective processing and encoding. This segmentation enables the system to leverage neural network potential for compression while managing implementation complexity by handling different parameter subsets separately through video-specific dictionaries

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces video-specific dictionaries as intermediary structures that bridge traditional compression frameworks and neural network capabilities. These dictionaries serve as mediators that enable neural network-based compression efficiency while maintaining compatibility with existing compression workflows and reducing implementation barriers

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4675498A1Video specific dictionary learning for implicit neural compression
Publication Date: 2026.01.07 INTERDIGITAL CE PATENT HOLDINGS SAS
  • EP4675498A1 patent drawingFigure 1
  • EP4675498A1 patent drawingFigure 2
  • EP4675498A1 patent drawingFigure 3

AI summary

Methods and apparatus are provided for encoding and subsequent decoding of video data by using a learnt video specific dictionary for implicit neural compression. An implicit neural representation comprising a head layer and a tail layer is used with approximations to the head layer parameters. The approximations are determined with combinations of atoms of a learnt video specific dictionary. In one embodiment, the head layer approximations and tail layer parameters are encoded in a bitstream. The dictionary is learnt at the decoder. In another embodiment, a dictionary is sent in the bitstream. At decoding, a reconstructed image is computed using transmitted INR parameters.