融入残差注意力机制的DeepLabV3+图像拼接篡改取证网络  

DeepLabV3+image splicing tampering forensic network fused residual attention mechanism

作　　者：吴云 张玉金 江潇潇[1] 许灵龙 WU Yun;ZHANG Yujin;JIANG Xiaoxiao;XU Linglong(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《光电子．激光》2023年第9期923-931,共9页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61902237);上海市自然科学基金项目(17ZR1411900)。

摘　　要：针对现有图像拼接检测网络模型存在边缘信息关注度不够、像素级精准定位效果不够好等问题,提出一种融入残差注意力机制的DeepLabV3+图像拼接篡改取证方法,该方法利用编-解码结构实现像素级图像的拼接篡改定位。在编码阶段,将高效注意力模块融入ResNet101的残差模块中,通过残差模块的堆叠以减小不重要的特征比重,凸显拼接篡改痕迹;其次,利用带有空洞卷积的空间金字塔池化模块进行多尺度特征提取,将得到的特征图进行拼接后通过空间和通道注意力机制进行语义信息建模。在解码阶段,通过融合多尺度的浅层和深层图像特征提升图像的拼接伪造区域的定位精度。实验结果表明,在CASIA 1.0、COLUMBIA和CARVALHO数据集上的拼接篡改定位精度分别达到了0.761、0.742和0.745,所提方法的图像拼接伪造区域定位性能优于一些现有的方法,同时该方法对JPEG压缩也具有更好的鲁棒性。Aiming at the problem that the existing image splicing detection network model has insufficient attention to edge information and is not good enough for pixel-level accurate localization effects,a DeepLabV3+image splicing tampering forensic method incorporating a residual attention mechanism is proposed.The methods use an encoding-decoding structure to achieve pixel level image splicing tampering localization.In the coding stage,the efficient attention module is integrated into the residual module of ResNet101.The residual module is stacked to reduce the proportion of unimportant features and highlight the splicing tampering traces.Then,the spatial pyramid pooling module with hole convolution is used for multi-scale feature extraction.The obtained feature maps are stitched and then modelled by spatial and channel attention mechanisms for semantic information.In the decoding stage,the localization accuracy of the image splicing forgery region is improved by fusing multi-scale shallow and deep image features.The experimental results show that the localization accuracy of splicing tampering on CASIA 1.0,COLUMBIA and CARVALHO datasets reaches 0.761,0.742 and 0.745,respectively.The proposed method has better image splicing forgery region localization performance than some existing methods,and the network also has better robustness to JPEG compression.

关 键 词：数字图像取证 图像拼接检测 改进的DeepLabV3+网络 高效通道注意力模块(ECA) 深度可分离卷积 

分 类 号：TP309.2[自动化与计算机技术—计算机系统结构]