Research On Key Theories And Technology Of Digital Image Forensics Based On Deep Learning

With the popularity of figure-revision software,it is easier to edit and revise photos,which breeds a hotbed of tampered images.The proliferation and dissemination of tampered images have brought many adverse effects and potential threats to social stability and judicial justice.Image forensics technology under blind environments and active protection methods represented by digital watermarking has achieved good research results,but the traditional methods often rely on experts'experience and professional knowledge of image to extract features.Deep learning has achieved amazing results in many fields such as image classification and recognition,but its application in the field of image forensics has just begun.Unique advantages of deep learning in automatic learning features make the research of key theory and technology in image forensics based on deep learning increasingly significant,which contributes to solving the challenges encountered by traditional forensic technology together with great research value.In this paper,we analyze the existing digital image tampering identification methods and propose theoretical innovation and algorithm improvement in several aspects of digital image tampering identification technology with the help of deep learning technology.The main achievements and innovations are as follows.1.In the aspect of asymmetric cropping tampering detection,we propose an image forensics technology based on parameter optimization of camera calibration.We use the imaging principle of camera calibration to solve the problem of image cropping tamper identification,and reduces the assumption of too many camera internal parameters.By retains four parameters of the original five parameters in the camera and through some approximation processing,we obtain the principal point through two regular geometric figures which are not coplanar in a single image.According to the offset of the principal point,the asymmetric cropping tampering behavior in the image can be identified.2.To deal with the location of image tampered region,we propose a tamper identification technology based on the full convolutional network.By making use of the segmentation ability of full convolution network,we apply the region segmentation ability to image forensics.On the basis of changing the classification of full convolution network,it is used to locate the tampered region.We first add the original image to the training data set containing the tampered image to form a positive sample and a negative sample,while setting the label image of the original image as a black image.The purpose of forming a positive sample and a negative sample is to guide the improved FCN to distinguish the difference between the original image and the tampered image.Finally,we conduct experiments to validate our algorithm,and the results show that the improved FCN can indeed achieve accurate regional positioning.3.In the aspect of image splicing location detection,we propose a correlation model based on DeepLabv3+and SRM.The original DeepLabv3+network is modified by classification number to learn the contour of spliced region,and the tag of different areas in spliced image is used to make it able to learn the edge features of spliced region.In addition,we obtaine the noise map by passing the input image to the SRM filter layer and provide additional evidence for splicing operations using noise inconsistencies between real and spliced regions.Then,through the correlation analysis of DeepLabv3+and SRM,we realize the precise location analysis of the spliced region.4.In the identification of face images,face feature detection algorithms based on deep learning are proposed in view of the fact that some face occlusion and generated faces by AI are becoming increasingly common.First,we use CNN network to extract features of face images,and automatically select the similar image sequence by the improved sparse expression method for feature matching,which effectively utilizes the related information between image sequences,eliminates some sub-blocks that affect the matching effect,and improves the accuracy of matching.Next,we propose an image identification method based on preloaded VGG network to distinguish AI generated face images from original images and make a preliminary exploration of AI technology to combat AI fraud.