Research On Detection Of Image Region Forgery Affected By Natural Interference

With the advent of the information age, the digital image has spread to people s dailylives. At the same time, all kinds of image editing softwares are steadily on the increase,which have posed a severe negative impact on image information security in different imageapplication areas. Image region duplication forgery tampers in one image for copying certainkey areas and pasting them. Today, natural interference and man-made interference havebecome two key issues facing image region duplication forgery. So far, region duplicationforgery detection existing man-made interference has made a certain achievement, however,natural interference, such as scarcely textured region and intrinsically identical objects hasbeen the difficulty for the tamper detection. They are often mistaken for duplication forgeryregion, which will reduce detection accuracy.In view of the natural interference can cause false positive when the image is detectedby region-duplication forgery detection. For the scarcely textured region, a recognitionalgorithm is proposed based on fractal dimension and image entropy. Fractal dimension isestimated by differential box counting and information is estimated by image entropy of theimage mapped to the3-D space, the combination of these two technologies can extractimage texture feature effectively, then the features are inputted to the support vector machine（SVM）.For the intrinsically identical objects, a detection algorithm is proposed based on localoutlier estimation and cluster compactness. Firstly, make the second judgment for theduplication region and the forgery region after the image detection. According to thedifferent distribution character of the feature points for the intrinsically identical objects andthe duplication forgery region, SIFT feature is extracted of the suspicious area pair andbidirectional matching is applied to improve the matching accuracy. Secondly, normalize thedistance of matching points, then local outlier factor and local reachability density areapplied to analyze distribution of matching points after estimating influence on itsneighborhood by counterfeit-Gaussian influence function in order to estimate the distributionof matching points accurately. Thirdly, due to some image area are often has few details,feature points reflect local compactness, Affinity Propagation is needed in the process ofestimating cluster compactness. Finally, local outlier factor and average cluster as the feature vector input into the SVM for training and through the optimal discriminant plane todistinguish the intrinsically identical objectsExperimental results show that the method above, can resist detecting false positivecaused by scarcely textured region and intrinsically identical objects and can resist attackwhich include gauss white noise and JPEG compression effectively. The experimental resultsalso prove that the method has high accuracy rate and low false detection rate.