Research On Passive Forensics Methods Based On Noise Feature For Image Splicing/Composite Detection

With the rapid development of digital information technology,image acquisition devices such as digital cameras and smart phones are widely used,and digital image can be obtained everywhere.Meanwhile,the content of digital image can easily be changed without leaving obvious tampering trace due to the increasing availability of free image editing software.The existence of tampered images not only reduces the credibility of image information,but also changes the traditional concept of"seeing is believing".Therefore,the detection techniques that reveal various image tampering traces have been paid more attention in recent years.Image splicing/compositing is a most common content tampering operation.In this work,we dedicate to the study of the detection method for image splicing/compositing tampering.The main works are follows:We proposed an image splicing detection method based on the noise distribution characteristics in different image regions.In the proposed method,we use SLIC algorithm to segment the test image into irregular super pixel blocks.Then we utilize the fact that image blocks from different sources tend to have larger inter-class difference,and luse PCA-based noise estimation method and Poisson-distribution-based noise probability estimation method,combine with the fuzzy c-means clustering to identify spliced/composited image regions.The proposed method can identify the splicing/compositing image regions,the detection accuracy reaches pixel level,and it is robust for content-preserving manipulations.Comparing with existing noise-based image splicing region detection methods,the proposed method has superior performance,especially when the noise difference between the spliced region and the original region is small.We proposed an image splicing region detection method based on the noise distribution level inconsistency.In this work,we use the local gradient of the non-overlapping image blocks as the texture feature,and use the fuzzy c-means clustering algorithm to estimate threshold value to determine the suspicious image region preliminarily.Then we use the characteristic that Laplace operator can enhancement the effect of the noise,and combine the singular value decomposition to extract image noise of suspicious region.According to the inconsistency of noise level to located spliced image regions.The proposed method is robust to content-preserving manipulations such as JPEG compression,Gaussian blur,gamma correction,down sampling and so on.Considering that most digital cameras generate CFA interpolation operation through the sensor during the imaging process,the photos taken by the same camera will have consistent interpolation mode.If the splicing regions of an image come from different cameras,interpolation artifacts will be caused by inconsistency of interpolation modes.We used CFA artifacts to estimate the image noise residual,and then use the inconsistency of the noise residual to detect spliced/composited image regions.In the proposed method,on the one hand,we use the double-tree complex wavelet transform and Wiener filter to process the green component of the image,obtain the noise residual,and calculate the variance of local noise residual.On the other hand,we use CFA interpolation method to predict the green component of the image,and estimate the noise residual of the green channel,then calculate the variance of local noise residual.Then we use above two variances to construct the forensic features,and use the OSTU threshold selection algorithm to locate the spliced/composited image regions.Experimental results show that this method has good detection performance for image splicing and tampering,and its pixel-level tampering detection accuracy reaches 90.5%.Compared to the existing image splicing region detection methods,the proposed method has superior performance.However,the limitation is that it is only effective for detecting spliced/composited image that come from different sources.