尺度可变有约束图像拼接检测与定位及其对抗优化  

Scalable constrained image splicing detection and localization with adversarial optimizing

作　　者：刘亚奇 蔡强[2] 石磊[3] 张一凡 吕斌斌 夏超 许盛伟[1] LIU Yaqi;CAI Qiang;SHI Lei;ZHANG Yifan;LYU Binbin;XIA Chao;XU Shengwei(Beijing Electronic Science and Technology Institute,Beijing 100070,China;Beijing Key Laboratory of Big Data Technology for Food Safety,Beijing Technology and Business University,Beijing 100048,China;State Key Laboratory of Media Convergence and Communiaction,Communication University of China,Beijing 100024,China)

机构地区：[1]北京电子科技学院,北京100070 [2]北京工商大学食品安全大数据技术北京市重点实验室,北京100048 [3]中国传媒大学媒体融合与传播国家重点实验室,北京100024

出　　处：《智能系统学报》2024年第6期1479-1491,共13页CAAI Transactions on Intelligent Systems

基　　金：中央高校基本科研业务费资金项目(3282023016);国家自然科学基金项目(62102010,62002003);北京工商大学食品安全大数据技术北京市重点实验室开放课题(BTBD-2022KF02).

摘　　要：针对有约束图像拼接检测与定位这一图像取证任务,提出一种尺度可变的检测与定位方法及其对抗优化架构。在所提有约束图像拼接检测与定位网络中,设计基于高效通道注意力增强的尺度可变关联性计算模型,采用高效通道注意力增强对通道特征进行校准,利用强关联因子截断操作实现处理任意尺度大小的图像;设计一种空间注意力增强的空洞空间金字塔池化对多尺度信息进行挖掘,以及基于深度可分离卷积和残差结构的定位结果重构网络。此外,提出一种块级对抗学习机制对预训练的尺度可变有约束图像拼接检测与定位网络进行优化。在公开数据集上的大量实验证明所提方法的有效性。A scalable detection and localization method,along with its adversarial optimization architecture,is proposed for the image forensics task of constrained image splicing detection and localization(CISDL).In the CISDL network,a novel scalable correlation computation module is constructed using high-efficiency channel attention enhancement blocks.The channel features are then calibrated using high-efficiency channel attention enhancement.Images of arbitrary sizes are processed using truncation operations on closely correlated factors,and a mask reconstruction network is designed based on depthwise separable convolution and residual connections.Finally,a patch-level adversarial learning strategy is proposed to optimize the pretrained model.Extensive experiments on publicly available datasets demonstrate the effectiveness of the proposed method.

关 键 词：有约束图像拼接检测与定位 尺度可变 关联性计算 对抗学习 图像取证 空洞卷积 金字塔池化 深度可分离卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]