Super-Resolution Model Training With Masked Pixel Blocks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing super-resolution image processing technologies face challenges in learning information outside the receptive field, leading to degraded performance due to non-learning up-sampling processes.

Innovation Solution

A machine learning model is trained by expanding image pixels into pixel blocks, using masks to control information flow, and employing self-attention mechanisms to predict pixel values, thereby integrating feature extraction and up-sampling, while limiting irrelevant information learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If up-sampling technology is used for super-resolution processing, then the resolution of original images is improved, but the performance is degraded due to non-learning up-sampling processes that cannot learn information outside the receptive field

Engineering Contradiction:

Improveimage resolutionVSAvoidsuper-resolution processing performance

Core Design Contradiction:

Manufacturing precisionVSReliability

Solution Approach 1:

The image is divided into multiple pixel blocks, where each block contains pixels to be predicted and reference pixels. This segmentation allows the model to process local regions independently while maintaining global context through the self-attention mechanism, resolving the contradiction by enabling learning beyond the traditional receptive field limits.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimensional structure by organizing pixels into blocks with reference pixels positioned at specific locations. This block-based approach with spatially distributed reference pixels extends the effective receptive field in a novel dimensional arrangement, allowing the model to capture long-range dependencies while maintaining local detail accuracy.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Reliability

If pixel blocks with reference pixels are used to enable learning beyond receptive field, then super-resolution performance is improved, but computational resources increase

Engineering Contradiction:

Improvesuper-resolution processing performanceVSAvoidcomputational resources

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

Different pixels within the block serve different functions: reference pixels provide contextual information while pixels to be predicted focus on local detail reconstruction. This local quality differentiation allows the model to allocate computational resources efficiently, with reference pixels guiding the overall structure and predicted pixels focusing on fine details, thus improving performance without proportionally increasing computational burden.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

Reference pixels are positioned and prepared in advance within each block to provide guiding information before the prediction process. This preliminary arrangement of reference pixels allows the self-attention mechanism to efficiently compute correlations without examining the entire image, reducing computational resources while maintaining improved performance.

Inventive Principle:

Principle #10Preliminary action

3Loss of information

If self-attention mechanism is applied to pixel blocks, then the model can learn relevant information effectively, but the device complexity increases

Engineering Contradiction:

Improveinformation learning capabilityVSAvoidmodel structure complexity

Core Design Contradiction:

Loss of informationVSDevice complexity

Solution Approach 1:

The self-attention mechanism is applied locally within segmented pixel blocks rather than across the entire image. This segmentation reduces the complexity of the attention computation from O(N²) for the whole image to O(n²) for each small block, where n is the block size. The model thus learns relevant information effectively within each block while the overall device complexity remains manageable.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The self-attention mechanism computes correlations only for relevant pixel pairs within the block (pixels to be predicted and reference pixels) rather than all possible pairs. This partial action approach focuses computational effort on informative relationships, reducing device complexity while maintaining effective information learning capability.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS12518392B2Training method and apparatus for machine learning model, image processing method and apparatus

Publication Date: 2026.01.06 BEIJING BOE TECH DEV CO LTD

AI summary

The present disclosure relates to a training method and training apparatus for a machine learning model, and a method and apparatus for image processing, which relates to the technical field of image processing. The training method for a machine learning model includes: expanding at least one pixel of an image sample to be processed into a pixel block, which comprises a plurality of pixels to be predicted; processing the pixel block using a first mask and obtain a mask processed result; according to the mask processed result, predicting prediction pixel values of the plurality of pixels to be predicted using a machine learning model; training a machine learning model according to the prediction pixel values of the plurality of pixels to be predicted and labeled pixel values of the plurality of pixels to be predicted.