Facial Image Forgery Detection With Frequency-Spatial Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial image detection models based on facial forgery traces have limited generalization and accuracy, especially with the advancement of face-changing technologies that generate biologically consistent forged faces.
Innovation Solution
A method involving frequency-domain transformation and spatial-domain feature extraction, followed by fusion using an attention fusion network to enhance the detection accuracy of forged facial images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If face-changing technology is used to generate forged faces with biological patterns, then the realism of forged images is improved, but the detection accuracy of existing models deteriorates
Solution Approach 1:
The patent transforms the detection approach from spatial domain to frequency domain by applying Fourier transform. This dimensional change allows the model to detect forged images through frequency characteristics that are not visible in the spatial domain, effectively addressing forged faces with realistic biological patterns that evade spatial-domain detection methods
Solution Approach 2:
The patent combines spatial-domain features and frequency-domain features to create a composite feature representation. This composite approach integrates the strengths of both domains: spatial-domain features capture structural information while frequency-domain features reveal forgery artifacts, together providing robust detection accuracy across different forgery types
2Measurement precision
If detection models focus on specific forged traces like blink patterns, then detection performance on those specific traces is improved, but generalization to other forgery types deteriorates
Solution Approach 1:
The patent creates a universal detection framework that works across multiple forgery types by extracting features from both spatial and frequency domains. This multi-functional approach enables the single model to detect various forgery methods (face swapping, deepfakes, video manipulation) without requiring separate specialized detectors for each forgery type
3Measurement precision
If forged faces have biologically consistent patterns, then the realism of forged images is improved, but the detectability of forged traces deteriorates
Solution Approach 1:
The patent applies Fourier transform to convert images from spatial domain to frequency domain, revealing hidden forgery artifacts that are imperceptible in the spatial domain. This dimensional transformation makes detectable the subtle inconsistencies in forged images that have realistic biological patterns, by analyzing frequency characteristics rather than visual patterns
Data Source
AI summary
An image detection method includes obtaining a facial image, and obtaining a frequency-domain image of the facial image and a spatial-domain feature of the facial image, the frequency-domain image being obtained by performing frequency-domain transformation on the facial image. The spatial-domain feature is obtained by performing feature extraction on the facial image. The method further includes performing feature extraction based on the frequency-domain image, to obtain a frequency-domain feature of the facial image, and fusing the spatial-domain feature and the frequency-domain feature by using an attention fusion network of a facial image detection model, to obtain a fused feature of the facial image. The method further includes obtaining a detection result of the facial image based on the fused feature, the detection result indicating whether the facial image is a forged facial image.