计算机科学
探测器
人工智能
算法
噪音(视频)
度量(数据仓库)
模式识别(心理学)
钥匙(锁)
数学
数据挖掘
选择(遗传算法)
作者
Jialei Zhou,Hongnan Ma,Wenlong Wang,Shuxia Ren,Kexue Fu
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-21
卷期号:: 14122-14126
标识
DOI:10.1109/icassp55912.2026.11462533
摘要
With the rapid development of diffusion-based image generation, distinguishing highly realistic synthetic images from authentic ones has become increasingly crucial. Despite their visual realism, generative models inevitably introduce inconsistencies due to their fundamentally imitative nature. To address this, we propose UAFD (Unified Adaptive Frequency-Domain Detector), a detection framework that maps nuance, illumination, and reconstruction inconsistencies into a unified frequency-domain representation space. Specifically, nuance inconsistencies are captured via multi-scale Haar wavelet decomposition across different subbands; illumination inconsistencies are modeled through gradient analysis combined with multi-directional Gabor filtering to extract directional lighting cues; and reconstruction inconsistencies are redefined as residual differences between the original and diffusion-reconstructed image spectrums. Furthermore, a spectrum-adaptive weighting mechanism dynamically adjusts the importance of each modality based on frequency energy distribution and spectral complexity. This unified design enhances robustness across generative methods while improving generalization, interpretability, and efficiency.
科研通智能强力驱动
Strongly Powered by AbleSci AI