Gradient-Weighted Intra Prediction for High-Resolution Image Coding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image compression technologies struggle to effectively predict pixel values, particularly in high-resolution and high-quality images, leading to inefficiencies in compression and potential artifacts.

Innovation Solution

An intra prediction method that generates a final prediction sample for a current block based on a weighted sum of prediction samples from neighboring blocks, with weights determined by gradients calculated from the neighboring regions, considering both horizontal and vertical directions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If simple intra prediction methods are used, then device complexity is reduced, but prediction precision deteriorates for high-resolution images

Engineering Contradiction:
Improveprediction precisionVSAvoiddevice complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies different weight values to different prediction samples based on local gradient characteristics. By calculating gradients in horizontal and vertical directions from neighboring blocks and selecting appropriate weights (e.g., larger weight for horizontal prediction when horizontal gradient is smaller), the method adapts the prediction quality to local image characteristics, improving prediction precision without requiring a completely complex system

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent dynamically determines weight values based on gradient calculations from neighboring blocks. Instead of using fixed weights, the system calculates gradients in horizontal and vertical directions and adapts the weights accordingly (e.g., when horizontal gradient < vertical gradient, use larger horizontal weight). This dynamic adaptation allows the system to achieve high prediction precision for high-resolution images while maintaining manageable complexity through automated weight selection

Inventive Principle:
Principle #15Dynamics

2Measurement precision

If weighted sum of multiple prediction samples is used, then prediction precision is improved, but device complexity increases

Engineering Contradiction:
Improveprediction precisionVSAvoiddevice complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent changes the parameter of weight values based on gradient characteristics. By calculating gradients in horizontal and vertical directions from neighboring blocks and selecting appropriate weights (e.g., larger weight for horizontal prediction when horizontal gradient is smaller), the method adapts the prediction quality to local image characteristics, improving prediction precision without requiring a completely complex system

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The system uses information from neighboring blocks (gradients) to automatically determine appropriate weights for prediction samples. The gradient calculations and weight selection process is self-contained within the prediction mechanism itself, allowing the system to improve its own prediction precision without external intervention or overly complex external control systems

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4651492A1Image encoding/decoding method and apparatus, and recording medium having bitstream stored therein
Publication Date: 2025.11.19 LG ELECTRONICS INC
  • EP4651492A1 patent drawingFigure 1
  • EP4651492A1 patent drawingFigure 2
  • EP4651492A1 patent drawingFigure 3

AI summary

An image decoding/encoding method and apparatus according to the present disclosure may: on the basis of a left reference sample of the current block, generate a first prediction sample for the current sample within the current block; generate a second prediction sample for the current sample on the basis of an upper reference sample of the current block; and generate a final prediction sample for the current sample on the basis of the weighted sum of the first prediction sample and the second prediction sample. Here, weights for the weighted sum comprises a first weight applied to the first prediction sample and a second weight applied to the second prediction sample, and the first weight may be determined on the basis of one or more gradients calculated from a neighboring region of the current block.