摘要
为防止对图像内容进行操作(如拼接),提出一种基于注意力卷积神经网络的图像篡改定位算法,称为CA-Net。尽管卷积神经网络强大的特征学习和映射能力可依次获取丰富的空间特征,但是本文提出使用不同采样步长的并行孔洞卷积层以提取多尺度特征。同时,为了更好地利用特征通道信息,在解码网中额外引入通道注意力模块。实验采用Synthesized图像库进行训练,在两组图像库上NC2016和CASIA进行微调和测试。实验结果表明:提出的并行孔洞卷积层和通道注意力模块能明显提高篡改定位结果,与一些最新主流算法相比,CANet在两个标准图像库上表现最优。
To prevent manipulation of image content(such as splicing),this paper proposes an image manipulation localization algorithm based on Channel Attention Convolutional Neural Network,and it is called CA-Net. Although the powerful feature learning and mapping capabilities of CNNs can sequentially acquire rich spatial features,this paper proposes to use parallel dilated convolutional layers with different sampling steps to extract multi-scale features. At the same time,in order to make better use of the characteristic channel information,we additionally introduce a channel attention module in the decoding network. This experiment uses Synthesized image dataset for training,and fine-tunes and tests NC2016 and CASIA on the two image libraries. Experimental results show that the proposed parallel dilated convolutional layer and channel attention module can significantly improve the results. Compared with some of the state-of-the-art algorithms,CA-Net performs best on the two standard image datasets.
作者
钟辉
康恒
吕颖达
李振建
李红
欧阳若川
ZHONG Hui;KANG Heng;LYU Ying-da;LI Zhen-jian;LI Hong;OUYANG Ruo-chuan(Management Center of Big Data and Network,Jilin University,Changchun 130012,China;ZICT Technology Co.,Ltd.,Shenzhen 518000,China;Center for Computer Fundamental Education,Jilin University,Changchun 130012,China)
出处
《吉林大学学报(工学版)》
EI
CAS
CSCD
北大核心
2021年第5期1838-1844,共7页
Journal of Jilin University:Engineering and Technology Edition
基金
吉林省省级产业创新专项项目(2017C031-4)
赛尔网络下一代互联网技术创新项目(NGII20180104,NGII20181202).
关键词
通道注意力
卷积神经网络
图像篡改定位
特征提取
channel attention
convolutional neural networks
image manipulation localization
feature extraction