Generic Image Quality Assessment Model for Face Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current face recognition technologies face challenges in handling images captured under non-ideal conditions, such as low resolution, blur, large pose variations, and lighting changes, which affect the quality and consistency of images, making it difficult to match faces across varying qualities.

Innovation Solution

A computer-implemented method that assesses image quality using a generic model, reducing the need for input parameters and element detection, allowing for real-time processing even with low-quality images. This method constructs a model from multiple images, uses lower frequency components, and employs a Gaussian probability density function to determine similarity, enabling efficient computation and quality scoring without requiring separate algorithms for each aspect of image quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If current face recognition approaches use reference gallery comprising images of known faces and perform series of comparisons to actual reference images, then matching accuracy can be maintained across different quality images, but computation time increases and real-time processing becomes difficult

Engineering Contradiction:
Improvematching accuracyVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent pre-computes quality metrics and organizes reference images into quality-based groups before actual recognition occurs. By performing quality assessment and image organization in advance, the system reduces computation time during real-time processing while maintaining matching accuracy through pre-validated reference selections.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the reference gallery into multiple quality levels or groups based on pre-assessed image quality metrics. This segmentation allows the system to quickly select appropriate reference groups for comparison without examining every reference image, thereby reducing computation time while maintaining accuracy by matching probe images against appropriately quality-matched references.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If detection of actual elements of object such as eyes or nose is performed for quality assessment, then detailed quality metrics can be obtained, but the method becomes unsuitable for low quality images where element detection is impossible

Engineering Contradiction:
Improvequality metric detailVSAvoidsuitability for low quality images
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies partial action by using only the portions of images that are sufficient for quality assessment without requiring complete detection of all facial elements. By assessing quality based on available observable features rather than requiring full element detection, the method remains effective even when images are too low quality for complete feature detection.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

Instead of detecting facial elements and then assessing quality based on detection results, the patent inverts the approach by first assessing overall image quality metrics (such as blur, lighting, resolution) and then using those metrics to determine appropriate processing. This inversion allows the system to handle low-quality images by adjusting processing based on pre-assessed quality rather than requiring successful element detection first.

Inventive Principle:
Principle #13The other way round (Inversion)

3Adaptability or versatility

If multiple training images with different preferred characteristics are used to construct the model, then different quality assessments can be determined for different pre-processing purposes, but the model complexity and computation time increase

Engineering Contradiction:
Improvedifferent quality assessments for different purposesVSAvoidmodel complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent creates a universal quality assessment model that can serve multiple pre-processing purposes through a single integrated framework. By designing the model to handle various quality metrics (blur, lighting, resolution, pose) within one unified structure, it provides adaptability for different purposes without requiring separate models for each function, thereby reducing overall complexity.

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

Solution Approach 2:

The patent merges multiple quality assessment criteria and training image characteristics into a single integrated model. By combining different quality metrics and training approaches into one unified model rather than maintaining separate models for each quality aspect, the system achieves versatility for different pre-processing purposes while reducing computational overhead and model complexity through consolidation.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP2676224B1Image quality assessment
Publication Date: 2021.05.26 IOMNISCIENT
  • EP2676224B1 patent drawingFigure 1
  • EP2676224B1 patent drawingFigure 2~3
  • EP2676224B1 patent drawingFigure 4

AI summary

This disclosure concerns image quality assessment. In particular, there is described a computer implemented method, software, and computer for assessing the quality of an image. For example but not limited to, quality of the image of a face indicates the suitability of the image for use in face recognition. The invention comprises determining (112) a similarity of features of two or more sub-images of the image (608) to a model (412) of the object which is based on multiple training images (612) of multiple different objects of that type. The model (412) is comprised of sub-models (406) and each sub-model (406) corresponds to a sub-image of the image (608). Determining similarity is based on the similarity of features of each sub-image to features modelled by the corresponding sub-model. It is an advantage that no input parameters are required for quality assessment since the quality of the image can be determined from only the similarity between the image and the same, therefore single generic, model.