Learning Local Texture and Global Frequency Clues for Face Forgery Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Image Forensics

Background:

• Deep learning has advanced face forgery creation and detection.
• Existing face forgery detection methods lack generalization across datasets and techniques.

Purpose of the Study:

• To enhance the robustness and generalization of face forgery detection.
• To develop a method leveraging both local texture and global frequency domain information.

Main Methods:

• A local texture mining and enhancement module using image patching, masking, and texture enhancement.
• Multi-scale frequency domain feature extraction via wavelet transform.
• An innovative frequency-domain processing strategy with selection and dynamic weighting.
• An integrated framework combining texture and frequency features with spatial and channel attention mechanisms.

Main Results:

• The proposed method demonstrates superior performance on benchmark datasets.
• The technique shows improved generalization capabilities compared to existing methods.
• The combined approach effectively captures subtle forgery traces and frequency inconsistencies.

Conclusions:

• The integrated framework effectively leverages complementary local and global features for robust face forgery detection.
• The method offers improved generalization, addressing a key limitation in current detection techniques.