Face Forgery Detection via Texture and Saliency Enhancement

In recent years, AI-driven advancements have resulted in increasingly sophisticated face forgery techniques, posing a challenge in distinguishing genuine images from manipulated ones. This presents significant societal and trust-related concerns. Most current methods approach this task as a binary classification problem via CNN, which potentially overlooks subtle forgery cues due to the down-sampling operations. This may lead to inadequate extraction of discriminative features. In this paper, we argue that forgery clues are hidden within facial textures and salient regions. Motivated by this insight, we propose the Texture and Saliency Enhancement Network (TSE-Net). TSE-Net contains two primary components: the Dynamic Texture Enhancement Module (DTEM) and the Salient Region Attention Module (SRAM). DTEM employs the Gray-Level Co-occurrence Matrix to extract facial texture information, while SRAM concentrates on salient regions. Extensive experiments demonstrate that the TSE-Net has superior performance in comparison to competing methods, emphasizing its effectiveness in detecting subtle face forgery cues through enhanced texture and saliency information extraction.