Image Super-Resolution Training With Balanced Texture Sampling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing super-resolution methods based on deep neural networks suffer from a significant imbalance in the proportion of sampled samples, particularly due to the prevalence of smooth regions, leading to networks favoring smooth regions over textured areas, which results in a loss of clear details in the reconstruction process.

Innovation Solution

The method employs a pyramidal neural network structure that iteratively trains a target neural network model using a balanced sample set obtained by filtering original images based on intervals of sample information content, such as gradient or variance, ensuring equal sample distribution across intervals, and utilizes multi-norm regularization during training to enhance convergence efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If random sampling is used to obtain training samples, then the sampling process is simple and fast, but the proportion of smooth regions and textured regions becomes imbalanced, causing the network to favor smooth regions

Engineering Contradiction:
Improvesampling speedVSAvoidsample proportion balance
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The patent segments the training sample set into multiple intervals based on sample information content (such as gradient magnitude or variance). Samples are divided into intervals according to their characteristics, ensuring that both smooth regions and textured regions are represented proportionally in the training data, thereby resolving the imbalance caused by random sampling

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the sampling parameter from uniform random sampling to interval-based stratified sampling. By using sample information content (gradient, variance) as the basis for interval division, the sampling method adapts to the actual distribution of image regions, ensuring balanced representation of different texture types while maintaining efficient sampling

Inventive Principle:
Principle #35Parameter changes

2Ease of manufacture

If the network is trained with imbalanced samples, then the training process is straightforward, but the network loses detail information in textured regions during reconstruction

Engineering Contradiction:
Improvetraining process simplicityVSAvoidreconstruction detail quality
Core Design Contradiction:
Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent segments the sample space into multiple intervals based on sample information content, ensuring that textured regions and smooth regions are distributed evenly across different intervals. This segmentation allows the network to receive balanced training data that preserves detail information in textured regions while maintaining training simplicity

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a feedback mechanism where the network parameters are updated based on the reconstructed image quality and the original sample characteristics. The training process uses the interval-based sample distribution to guide parameter updates, ensuring that the network learns to preserve details in textured regions while maintaining overall reconstruction accuracy

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4152244B1Image super-resolution processing method, apparatus and device, and storage medium
Publication Date: 2026.02.25 SANECHIPS TECH CO LTD
  • EP4152244B1 patent drawingFigure 1
  • EP4152244B1 patent drawingFigure 1A
  • EP4152244B1 patent drawingFigure 1B~2

AI summary

Provided are an image super-resolution processing method and apparatus, a device, and a storage medium. The image super-resolution processing method includes acquiring an image to be processed and performing super-resolution processing on the to-be-processed image by a target neural network model. The target neural network model is obtained by iteratively training a pyramidal neural network model by using a target image sample set. The target image samples are obtained by filtering original image samples according to a plurality of intervals of a sample information content. The number of image samples in each interval of the sample information content is the same.