Digital Image Tampering Detection Based On Noise Characteristic Analysis

With the rapid development of intelligent digital image editing tools(such as Photoshop,Beauty Cam,etc.),digital image tampering is becoming easier and easier,and image splicing is one of the main tampering methods.Image splicing refers to combining two or more images to fake one image.Because the background and splicing areas of fake images are generally from different digital images,these images are usually taken by different imaging devices.The noise introduced by the sensors inside the imaging device often has certain differences,which makes the statistical characteristics of noise in the background and splicing areas have inconsistent.Therefore,the analysis of image noise characteristics is helpful for the detection and location of image tampering areas,and has high theoretical research and practical value.Existing image tampering detection algorithms based on noise characteristic analysis still have certain difficulties and challenges,which are mainly reflected in two aspects.On the one hand,when the noise difference between the background area and the tampered area is small,the existing methods are more difficult to distinguish between the tampered area and the background area;On the other hand,when there is a large difference in the texture of the image,the difficulty of noise analysis and estimation gradually increases,and the algorithm performance decreases significantly.This article conducts in-depth research on the above issues.The main research work is as follows:(1)Aiming at the noise inconsistency between the background area and the tampered area of the spliced forged image,this paper presents an image tampering detection algorithm based on statistical noise level inconsistency.When the noise difference between the background area and the tampered area is small,it is difficult for the existing methods to separate the tampered area from the background area.Based on this,a statistical noise level estimation algorithm is proposed.First,the image block noise level is estimated by calculating the eigenvalues in the redundant space dimension.Secondly,the clustering method is used to perform clustering based on the noise value of the image block.Finally,the tampered area is located through a two-stage strategy from coarse to fine.The experimental results show that compared with the existing algorithms,the image tampering detection method based on statistical noise level analysis proposed in this paper has significantly improved detection performance and is robust to different post-processing operations(such as JPEG compression,downsampling,etc.).(2)When there is a large difference in the texture of an image,it is increasingly difficult to use the above method to estimate the noise of an image block,resulting in a decrease in detection and localization performance.For this reason,this paper proposes image tampering detection based on low-rank image block noise analysis.In this paper,low-rank image blocks are selected using the local image gradient matrix and statistical characteristics.Use an iterative noise estimation algorithm for the selected low-rank image block to estimate the noise level of the entire image block.According to the image redundancy,the image block noise level estimation is more accurate.Finally,the tampered area is located by the image tampering positioning algorithm.Experimental results show that compared with the above detection algorithms,the detection performance of the proposed algorithm is further improved,and the tampered image can still locate the tampered area well after the tampered image undergoes JPEG compression and downsampling operations.The image tampering detection based on statistical noise level analysis is mainly aimed at the small difference between the background area and the tampered area of the tampered image,and the image tampering detection based on the low-rank image block noise analysis is not only effective when the noise difference between the background area and the tampered area is small,and it is effective when the difference in image texture changes is large,which further improves the detection performance and improves the shortcomings of the former method.