Abstract: Abstract: Existing Downsampling operation will make the image lose high-frequency forensic traces and detail information, increasing the difficulty of image forensics, and the existing deep learning-based image forensic network can not effectively detect the image tampered with by the downsampling operation, so to improve the robustness of the downsampling image forensics method becomes a bottleneck in image forensics. To solve this problem, a downsampling image forensic network HirrNet is proposed, which consists of two parts: an image recovery module and a forgery detection module. The image recovery module uses the idea of hierarchical conditional flow to reduce the loss of high-frequency information by recovering forged image forensic traces and detail information to improve the performance of forgery detection; the forgery detection module uses an end-to-end image segmentation network (RRU-Net) for tampering detection, which combines the Spatial Channel Attention (SCSE) mechanism to efficiently enhance the forgery-related features in the downsampled image Extraction.deep learning-based image forensic networks often cannot effectively detect tampered images that have undergone downsampling operations. The downsampling operation loses image high-frequency forensic traces and detail information, increasing the difficulty of image forensics, leading to the improvement of the robustness of the downsampled image forensics method as a bottleneck in image forensics. To solve this problem, this paper proposes a downsampling image forensic network HirrNet, which consists of two parts: an image recovery module and a forgery detection module. The image recovery module uses the idea of hierarchical conditional flow to improve the forgery detection performance by recovering the forged image forensic traces and detail information in order to reduce the loss of high frequency information. The forgery detection module performs tamper detection using an end-to-end image segmentation network (RRU-Net) that incorporates the spatial channel attention (SCSE) mechanism to enhance the extraction of forgery-related features in downsampled images. The experimental results show that HirrNet outperforms the comparison networks ForSim, PSCC, and DFCN in terms of F1 value, area under the curve (AUC), and intersection and merger ratio (IOU) on the DSO, Columbia, CASIA, and NIST16 datasets, and improves the accuracy by more than 10% and 14.5% for the tampered images whose sizes have been reduced to one-half and one-fourth of the original images .Experimental results show that HirrNet outperforms the comparison networks ForSim, PSCC, and DFCN in terms of F1 value, area under the curve (AUC), and intersection and merger ratio (IOU) on the DSO, Columbia, CASIA, and NIST16 datasets, and improves the accuracy by more than 10% and 14.5% for the downsampled tampered images with 2-fold and 4-fold reduction. The method in this paper can effectively solve the problem of poor robustness of existing downsampled image forensics methods.

Key words: image forensics, image recovery, spatial channel attention, RRU-Net, downsampling

摘要： 摘 要: 下采样操作会使图像丢失高频取证痕迹和细节信息，增加图像取证的难度，而现有的基于深度学习的图像取证网络也不能有效检测经过下采样操作篡改的图像，因此提高下采样图像取证方法的鲁棒性成了图像取证的瓶颈现有的基于深度学习的图像取证网络，往往不能有效地检测经过下采样操作的篡改图像。下采样操作会丢失图像高频取证痕迹和细节信息，增加图像取证的难度，导致提高下采样图像取证方法的鲁棒性成为图像取证的瓶颈。为解决这个问题，本文提出一个下采样图像取证网络HirrNet，该网络由两2部分组成：图像恢复模块和伪造检测模块。图像恢复模块使用分层条件流的思想，通过恢复伪造图像取证痕迹和细节信息，以减少高频信息的丢失，从而以提高伪造检测性能。；伪造检测模块则使用端到端图像分割网络(RRU-Net)进行篡改检测，该网络结合了空间通道注意力(SCSE)机制，可有效，以增强下采样图像中与伪造相关特征的提取。实验结果表明，HirrNet在 DSO、Columbia、CASIA和NIST16数据集上的F1值、曲线下面积(AUC)和交并比(IOU)优于对比网络ForSim、PSCC和DFCN，对于尺寸缩小到原图一半和2倍 和四分之一缩小4倍的篡改图像，准确率提高了10%和14.5%以上。HirrNet本文方法可以有效解决缓解现存的有下采样图像取证方法鲁棒性差的问题。

关键词: 图像取证, 图像恢复, 空间通道注意力, RRU-Net, 下采样