Research On HDR Imaging Forensics Technology Based On Machine Learning

As technology bring convenience,people's security is also threatened by technology.Nowadays,digital images are already a kind of information transmission medium that people can get at their fingertips,but the authenticity and integrity of images are often ignored.Image forensics is the subject that uses the statistical features of images to verify the authenticity,originality,and integrity of digital images.High Dynamic Range(HDR)image are known by its rich brightness range and excellent visual feeling.Unfortunately,few scholars are currently concerned about the safety of HDR images.Image forensics includes image source forensics,image tampering forensics and image steganographic analysis.Most of the existing image forensics algorithms only focus on LDR images and are not effective or even completely unsuitable when applied directly to HDR images.About image source forensics,this paper uses image joint histogram to analyze the difference in statistical features of HDR images generated by different methods,and then proposes a new concept of synthetic pattern noise for represent the inherent noise in the process of generating HDR images.Then we propose a neural network for HDR image source forensics based on the high spatial resolution,high gray resolution and unique compression style of HDR image,and achieves good results in the image brightness domain.Finally,the accuracy of the HDR image source forensics problem is further improved by the replacement of the residual structure.On the other hand,about the blurring operators detection,this paper analyzes the pixel value changes introduced by HDR images by different blurring operations through logarithmic luminance histogram,and quantizes the histogram as the statistical features of images under different blurring operations.In addition,this paper also proposes a new neural network for HDR image blurring operator detection to improve the initial feature extraction ability and fitting ability of the network.The experimental part of this paper makes some comparisons in the structure of neural networks,and also compared with traditional image forensics features and deep learning models.The results show that our method can achieve higher accuracy and robustness.