Differentiable Correlation Loss for Image Quality Ranking

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing data-driven approaches for image quality assessment (IQA) face challenges due to the high cost and labor-intensity of collecting large datasets, leading to models requiring excessive training data and being prone to overfitting, and the lack of effective methods for image content-agnostic quality comparisons.

Innovation Solution

A machine learning algorithm that employs differentiable correlation coefficients, such as Pearson and Spearman, to enforce monotonic relationships and rank correlations, allowing for image content-agnostic sampling and training with 'batch all' strategies that increase the training signal by comparing distorted images from different references, thereby improving model performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If large annotated datasets are collected for model training, then model training effectiveness is improved, but the cost and labor intensity of data collection increase significantly

Engineering Contradiction:
Improvemodel training effectivenessVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system uses automatic quality assessment algorithms to evaluate images without requiring manual human annotation. The model trains on automatically generated quality scores rather than hand-labeled data, eliminating the labor-intensive annotation process while maintaining training effectiveness

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

Manual human annotation processes are replaced with automated machine learning-based quality assessment systems. The mechanical process of human evaluators scoring images is substituted with computational algorithms that automatically generate quality metrics for training data

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

2Adaptability or versatility

If the number of model parameters is increased to improve expressive power, then model flexibility is improved, but the requirement for training data increases dramatically

Engineering Contradiction:
Improvemodel flexibilityVSAvoidtraining data quantity
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The system generates its own training data through automatic quality assessment of images in the dataset. This self-generated training data allows the model to learn from abundant examples without requiring external annotated datasets, enabling high-capacity models to train effectively on unannotated or lightly-annotated data

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system performs preliminary automatic quality assessment on a large corpus of images to generate training labels before model training begins. This pre-computed training data preparation enables subsequent training of complex models without requiring manual annotation during the training process

Inventive Principle:
Principle #10Preliminary action

3Quantity of substance

If image content-constrained comparisons are used for training, then training data availability is improved, but the ability to perform image-agnostic quality assessment is reduced

Engineering Contradiction:
Improvetraining data availabilityVSAvoidimage-agnostic assessment capability
Core Design Contradiction:
Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The system trains the model to perform quality assessment across diverse image contents and domains simultaneously. By using automatic assessment on varied datasets without content-specific constraints, the model learns universal quality patterns that generalize across different image types, enabling image-agnostic quality evaluation

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4309118B1Data agnostic sampling and differentiable correlation loss function
Publication Date: 2026.02.25 HUAWEI TECH CO LTD
  • EP4309118B1 patent drawingFigure 1
  • EP4309118B1 patent drawingFigure 2
  • EP4309118B1 patent drawingFigure 3~4

AI summary

Described is a processing mechanism (1000) for assessing a plurality of data units (201, 202), the processing mechanism comprising one or more processors (1001) configured to process the data units by means of a machine learning algorithm and thereby form, for each data unit, an output value representing both (i) a rating for that unit and (ii) a ranking of that unit relative to the other data units of the plurality of data units. This may allow a property of the data units, such as their quality, to be easily compared using the output value.