Research On HDR Image Forensics Based On Edge Difference

Compared with the traditional low dynamic range(LDR)imaging,high dynamic range(HDR)imaging can represent a larger dynamic range,restore the brightness and color of the scene better,and display more detailed information.However,the popularity of HDR imaging technology also brings new challenges to image security.The reason is that most image forensic algorithms currently focuse on LDR images only,and extending them to HDR images is not trivial.For example,forensic methods based on the statistical characteristics of natural images are often ineffective in HDR images due to the difference in statistical features between HDR and LDR.Traditional HDR detection algorithms have high requirements for image quality and content.Low-quality edges can cause the algorithm to fail or produce incorrect results.Therefore,researching the forensic problem of HDR images is of theoretical and practical importance.In the previous research on HDR image forensics,most studies believed that there were traces related to HDR forensics in the edge area.Although relevant researchers in the field of HDR image generation continuously optimize the generation algorithm,the generated traces still remain.For example,stripe-like halo effects caused by HDR image edges,noise appearing at the edges,or ghosting/blurring effects occurring at image edges.Therefore,based on edge traces,we further investigate HDR images authentication to solve technical problems in HDR image forensics.This thesis proposes an SAA-Net network based on a spatial rich model(SRM)edge enhancement block and the dual attention mechanism to distinguish between HDR images from LDR ones.Firstly,to preserve more subtle edge features,we propose a fully convolutional down-sampling module which serves as the basic architecture of the network and allows the convolutional block to gain a down-sampling approach that pays more attention to edge differences.Secondly,we embed an edge enhancement module that uses the SRM module to obtain various edge information.Finally,the use of spatial attention mechanisms and channel attention mechanisms at different scales can guide feature extraction networks at various levels to pay more attention to edge discrepancy information.In the field of image forgery detection research,besides identifying the authenticity of the image,splice tampering detection is also indispensable.Therefore,we apply the proposed HDR image forensics network to splicing localization.On multiple public databases and self-created databases,our proposed network demonstrates advanced performance in distinguishing HDR images,both real and synthesized,from LDR images.When applying HDR recognition to each part of a test image,we can accurately locate the tampered image composed of HDR and LDR images.In our self-created tampering database,the F1 value of splicing localization reaches 94.57%.