The Research Of Image Forensics Based On Subtle Features Analysis

Digital images,which serve as the important information carriers,have penetrate into our daily life.However,the rapid spread of illegal images and malicious manipulated images have had a major negative impact on the society,science and so on.Identifying the originality and authenticity of digital images has recently attracted considerable attention in the image forensic field of research.Image forensics is designed to determine the source of a digital image and whether it has undergone tampering by analyzing the changes in the statistical properties of digital images,as well as the "fingerprints" that persist at each stage of the imaging process Several algorithms have been established based on the hand-crafted features and deep learning.Although these approaches have led to a breakthrough in the image forensics,some important problems still remain unsolved.For instance,extracting the hand-crafted features with human efforts is a difficult and time-consuming process,while data-driven deep learning methods tend to learn features that represent image contents rather than cameras characteristics.To fully take advantages of both hand-crafted and data-driven technologies,we integrates the domain knowledge into deep learning methods to enhance the subtle features.Based on the above ideas,we propose a domain-knowledge driven multi-task learning camera identification method and a manipulated image detection and segmentation model based on multi-scale noise constraints.The domain knowledge-driven multi-task learning camera identification method consists of one pre-processing module,one feature extractor,and one hierarchical multi-task learning procedure.The backbone of the feature extractor is ResNet which can integrate the low/mid/high-level features together.The pre-processing module can suppress the influence of the im-age contents as much as possible,and introduce the domain knowledge to the subsequent deep learning network by the "split-transform-merge" training strategy.Moreover,for device-level identification,hierarchical multi-task learning can provide more supervise information from the brand and model.Our experimental results demonstrate that the proposed method is effective and robust.The accuracies of consumer-level camera identification in brand/model/device-level are 99.6%,97.1%and 52.4%,respectively.The accuracy of cell-phone device identification can even reach 84.3%.Our method is also robust to the manipulated images.Moreover,the t-SNE visualization results confirm that the features of different cell-phone devices are visually more separable than consumer-level cameras.Aiming at the problem of authenticity of image,we propose a manipulated image detec-tion and segmentation model based on multi-scale noise constraints.We use the multi-scale constrainted convolutional layer to extract the high-frequency noise residuals at different scales adaptively,and adopt the multi-task learning method to simultaneously classify,detect and seg-ment the manipulated images.The proposed method can reach 66.7%,42.5%and 89.8%F1 scores in COVER,CASIA,and NIST 16 datasets,respectively.The experimental results verify the effectiveness of the proposed method,and it can help improve the performance of detecting the manipulated images.