哈尔滨商业大学学报（自然科学版）2025，Vol.41Issue(6)：668-674,7.
融合双流特征的低光照伪造人脸检测方法
Low-light forged face detection method by integrating dual-stream features
摘要
Abstract
To address the poor performance of current facial forgery detection models under low-light conditions,a dual-stream network-based forgery detection model integrated with low-light enhancement technology was proposed.This model aimed to improve detection performance across various lighting conditions.Innovative low-light enhancement and directional denoising techniques were employed to amplify forgery-related features,while the discrete wavelet transform was utilized to extract frequency-domain information of images,facilitating the capture of detailed features at different scales.Subsequently,the backbone network learned both frequency-domain and spatial-domain features of the images,and these two types of features were fused.An attention mechanism was introduced to focus on the channel and spatial information of the fused features,with the final discrimination of image authenticity performed by a classifier.Evaluations conducted on in-dataset and cross-dataset scenarios demonstrated improvements in Accuracy(ACC)and Area Under the Curve(AUC).Additionally,ablation experiments verified the effectiveness of each module.The proposed model exhibited robust performance under varying lighting conditions,effectively capturing diverse forgery information and enhancing the accuracy and generalization ability of facial forgery detection.关键词
伪造人脸检测/双流网络/特征融合/低光照增强/离散小波变换/定向去噪Key words
fake face detection/two-stream network/feature fusion/low-light intensity/discrete wavelet transform/directional denoising分类
信息技术与安全科学引用本文复制引用
宋一帆,杨高明..融合双流特征的低光照伪造人脸检测方法[J].哈尔滨商业大学学报（自然科学版）,2025,41(6):668-674,7.基金项目
国家自然科学基金(No.52374155) （No.52374155）
安徽省自然科学基金(No.230805MF218) （No.230805MF218）
安徽理工大学2024年大学生创业基金扶持项目(淮南健芯医疗器械有限公司) （淮南健芯医疗器械有限公司）
