Facial Landmark Angular Comparison for Morphed Image Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Face recognition systems are vulnerable to morphed facial images, which can be manipulated to resemble multiple individuals, leading to incorrect authentication and verification, and existing detection methods struggle with accuracy, especially when images are printed and scanned, resulting in loss of detectable artifacts and reduced system performance.

Innovation Solution

A method that compares facial landmarks of a provided image with those of a captured, trusted image to determine if the provided image is artificial by calculating angular and distance values between corresponding landmarks, using a trained classifier to classify the image as real or artificial, without requiring structural analysis of the provided image.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing detection methods analyze structural features and patterns in facial images, then detection capability is provided, but detection accuracy is insufficient especially for printed and scanned images

Engineering Contradiction:
Improvedetection accuracyVSAvoidloss of detectable artifacts
Core Design Contradiction:
Measurement precisionVSDifficulty of detecting and measuring

Solution Approach 1:

The patent extracts only the essential facial landmark positions from facial images, separating this critical information from the rest of the image data. By focusing solely on landmark coordinates rather than analyzing entire image structures or patterns, the method avoids loss of detectable artifacts during printing and scanning while maintaining detection accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms facial image data into a different parameter space by converting image pixels into facial landmark coordinates. This parameter transformation from visual patterns to geometric positions makes the detection robust against printing and scanning artifacts, as landmark positions remain stable across these transformations.

Inventive Principle:
Principle #35Parameter changes

2Ease of operation

If facial images are printed and scanned for verification, then authentication process is completed, but detectable artifacts are lost reducing detection accuracy

Engineering Contradiction:
Improveauthentication processVSAvoiddetection accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent creates a simplified copy of the facial image in the form of facial landmark coordinates. This coordinate-based representation serves as a robust copy that can be reliably compared across different image formats (original, printed, scanned) without suffering from the artifacts that affect visual pattern analysis.

Inventive Principle:
Principle #26Copying

3Adaptability or versatility

If morphed facial images are used to attack face recognition systems, then authentication vulnerability increases, but detection of artificial images becomes necessary

Engineering Contradiction:
Improveresistance to morphing attacksVSAvoidauthentication reliability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent performs preliminary detection of facial landmarks and compares them against expected patterns before final authentication decisions are made. This preliminary analysis of landmark positions allows the system to identify morphed images early in the authentication process, preventing unreliable authentication before it can compromise system security.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3642756B1Detecting artificial facial images using facial landmarks
Publication Date: 2024.05.22 HOCHSCHULE DARMSTADT
  • EP3642756B1 patent drawingFigure 1
  • EP3642756B1 patent drawingFigure 2
  • EP3642756B1 patent drawingFigure 3

AI summary

A method and a device for classifying facial images are described. The method includes retrieving a first plurality of facial landmarks of a provided facial image associated with a subject, determining a corresponding second plurality of facial landmarks of a captured facial image of the subject, extracting at least one feature vector based on differences between at least some of the first plurality of facial landmarks and corresponding facial landmarks of the second plurality of facial landmarks, wherein at least some values of the at least one feature vector represent angular values of facial landmarks, and supplying the at least one feature vector to a classifier to classify the provided facial image as a real facial image of the subject or as an artificial facial image.