检索规则说明:AND代表“并且”;OR代表“或者”;NOT代表“不包含”;(注意必须大写,运算符两边需空一格)
检 索 范 例 :范例一: (K=图书馆学 OR K=情报学) AND A=范并思 范例二:J=计算机应用与软件 AND (U=C++ OR U=Basic) NOT M=Visual
作 者:宋家骏 刘桂雄[1] 黄家曦 张国才 SONG Jiajun;LIU Guixiong;HUANG Jiaxi;ZHANG Guocai(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,China)
机构地区:[1]华南理工大学机械与汽车工程学院,广东广州510640
出 处:《中国测试》2023年第9期37-45,共9页China Measurement & Test
基 金:广东省重点领域研发计划项目(2019B010154003)。
摘 要:伪造图像若被不当利用会带来严重负面影响,不同伪造图像生成方法导致伪造属性差异,使得研究统一图像伪造检测、定位方法具有很大挑战性。该文提出一种应用U-HRNet+SoftTripleLoss的HiFi-Net伪造图像检测方法,首先采用U-HRNet替代HiFi-Net特征提取网络,其网络结构促进学习图像深层特征以获取更高级的语义信息,增加多个阶段、融合通道以改善高分辨率特征;其次引入SoftTripleLoss模块,学习无约束采样的伪造属性特征嵌入表示以改善特征嵌入分布,从而更好地区分细粒度伪造属性,进而提高细粒度伪造图像分类准确率。实验表明,使用上述技术构建的检测模型像素级别总体评价指标AUC、F1分别为0.9928、0.9760,较原文献模型提高0.0025、0.0082;图像级别总体评价指标细粒度属性分类准确率Acc达98.05%,较原文献模型提高1.23%。If counterfeit images are improperly used,it will have serious negative effects,and different forged image generation methods lead to differences in forgery attributes,which makes it very challenging to study unified image forgery detection and localization methods.In this paper,a forged image detection and localization method based on HiFi-Net applied U-HRNet and SoftTripleLoss is proposed.Firstly,U-HRNet is used to replace feature extraction network of HiFi-Net,and its network structure promotes the learning of deep features of images to obtain more advanced semantic information,and its multiple stages and fusion channels improve high-resolution features.Secondly,the SoftTripleLoss module is introduced to learn the unconstrained sampling of forged attribute feature embedding representation to improve the feature embedding distribution,so as to distinguish fine-grained forged attributes better,and then improve the classification accuracy of finegrained forged images.Experimental results show that the overall pixel-level AUC and F1-scorce of the model mentioned above are 0.9928 and 0.9760,respectively,which are 0.0025 and 0.0082 higher than the original model.At the image level,the overall accuracy of fine-grained attribute classification reach 98.05%,which is 1.23%higher than that of the original model.
关 键 词:伪造图像 伪造属性分类 HiFi-Net U-HRNet SoftTripleLoss
分 类 号:TP391.4[自动化与计算机技术—计算机应用技术]
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在链接到云南高校图书馆文献保障联盟下载...
云南高校图书馆联盟文献共享服务平台 版权所有©
您的IP:216.73.216.200