Face Morphing Detection Using Spatial-Frequency Noise Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face authentication systems face challenges in accurately determining whether face image data is a morphing image, which can lead to fraudulent identity verification, and existing methods that use special sensors or deep learning techniques are costly or limited in effectiveness.
Innovation Solution
A method that utilizes noise removal and analysis of both spatial and frequency domain information to detect morphing images, using a computer-based system to generate difference image data and apply feature extraction techniques to improve determination accuracy without requiring special equipment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If special sensors or deep learning techniques are used to detect morphing images, then determination accuracy is improved, but device complexity and cost increase
Solution Approach 1:
The patent extracts and analyzes only the noise components from face images by generating difference images between original and denoised images. This extraction approach focuses computational resources on the most informative elements (noise patterns) rather than processing entire images, achieving high determination accuracy with simpler device complexity
Solution Approach 2:
The patent introduces difference images as an intermediary representation between original face images and morphing detection results. These difference images serve as a mediator that highlights noise components while suppressing legitimate face features, enabling accurate morphing detection without requiring complex deep learning models or special sensors
2Measurement precision
If deep learning techniques are used to detect morphing images, then determination accuracy is improved, but processing time increases
Solution Approach 1:
The patent extracts only noise components through simple denoising operations and difference image generation, avoiding the computationally intensive training and inference processes of deep learning models. This extraction strategy achieves morphing detection with significantly reduced processing time while maintaining high accuracy
Solution Approach 2:
The patent employs computationally inexpensive operations (denoising, difference calculation, frequency analysis) that can be executed quickly and discarded, replacing expensive deep learning models. These simple processing steps are sufficient for the specific task of morphing detection and can be performed rapidly without requiring complex model architectures
3Reliability
If noise removal processing is applied to face images, then signal quality is improved, but legitimate face features may be altered
Solution Approach 1:
The patent extracts only the noise components by computing the difference between original and denoised images, rather than directly modifying the original face image. This extraction approach isolates noise for analysis while preserving the integrity of legitimate face features in the original image
Solution Approach 2:
The patent uses difference images as an intermediary to represent noise components separately from legitimate face features. This intermediary representation allows noise analysis to proceed without altering or losing information from the original face image, maintaining both signal quality and feature integrity
Data Source
AI summary
In a determination method, a computer executes processing including: generating face image data from which noise is removed by a specific algorithm from face image data when the face image data is acquired; generating difference image data concerning difference between the face image data that has been acquired and the face image data that has been generated; determining whether or not the face image data that has been acquired is a composite image based on information included in the difference image data; and determining whether or not the face image data that has been acquired is a composite image based on information included in frequency data generated from the difference image data in a case where the face image data that has been acquired is not determined to be a composite image.