Research On Image Multi-region Splicing Detection Based On Noise Distribution

With the rapid development of intelligent imaging devices,digital image has become an important carrier of information transmission.Meanwhile,as the operation of image editing softwares becomes more and more humanized,people can modify the image content,which reduces the credibility of images.Therefore,the detection technology to judge the authenticity of digital image arises at the historic moment.Image splicing is a forgery method commonly used by tampers.Many scholars have proposed various of methods to localize the spliced region.However,with the increasing maturity of tampering technology,it is still necessary to explore the methods of image tampering detection.Aiming at the problems of low true positive rate and weak robustness of the current methods,a research on image multi-region splicing detection based on noise distribution is proposed.The concrete contents include:(1)To solve the problem of ignorance color information and the low accuracy of noise estimation for different size image blocks,a noise estimation method via adaptive quaternion singular value decomposition is proposed.First,the suspicious image is divided into super pixels.Then,the adaptive quaternion singular value decomposition is used to estimate noise of image blocks.Finally,the function between noise and brightness is constructed by calculating the brightness information of the super pixel to obtain the degree of constraint of super pixel blocks by the function curve which is used to describe the feature vectors of the noise.(2)Considering that images from different source may have similar noise level in some regions and single feature extraction methods cause weak robust for different scenes,it is proposed to use the camera's automatic white balance color temperature estimation algorithm to extract the color temperature features of super pixels.Then,the Hopkins statistic is used to evaluate the random distribution of noise and color temperature features to assign weights for both to obtain mixed features adaptively with good category.(3)Since the classification results of FCM clustering algorithm are strongly influenced by the initial clustering center.In addition,there are some small blocks in the super pixel segmentation results,bringing about a large possibility of such blocks being mislabeled in the detection regions.To solve above problems,the density and distance of feature vectors are used to study the optimal clustering centers,which can not only accelerate the convergence speed of FCM clustering algorithm,but improve the classification accuracy.Then,the initial suspicious regions are postprocessed with the context information of super pixel blocks to further improve the detection accuracy.Compared with other methods on the Columbia Uncompressed Image Splicing Detection Evaluation Dataset,the experimental results show that the proposed method in this paper improves the detection accuracy slightly.And it is robust for JPEG compression,Gaussian blur and gamma transform.Additionally,since there is no public multi-region image splicing dataset,the Columbia Uncompressed Image Splicing Dataset is used to construct multi-region splicing dataset in this work.The method in this paper is compared with the current method that can detect multi-region splicing images.The results show that the method of this paper on the self-constructed dataset has a certain effectiveness.