Research On Digital Images Forensics Of Diffusion-based Inpainting

Image inpainting is a kind of image processing technology,which is initially used to repair the damaged area and remove the stains and scratches in the image.From another perspective,the image inpainting technology is also a means of image tampering when it is used by the forger to remove the objects in the digital image.Nowadays,the increasingly mature image restoration technology can cover up the operation traces in the removal process.Under the increasingly popular image editing software and the rapid development of social networks,the emergence of a large number of forged images poses a serious threat to personal privacy and social stability.In recent years,image inpainting forensics has been concerned widely,which has high theoretical significance and practical value in the field of image forensics.Generally speaking,there are three categories of image inpainting technologies:exemplar-based,sparsity-based and diffusion-based technologies.At present,although several methods have been proposed for detecting exemplar-based and sparsity-based image inpainting,there is still a lack of effective method for detecting diffusion-based image inpainting.Therefore,this paper makes research on digital image forensics of diffusion-based inpainting,the main research work is as follows:(1)The diffusion-based image inpainting technology brings fuzzy effect to the inpainted areas in the diffusion process,which results in some texture changes in the inpainted areas.By analyzing the diffusion process of the image inpainting,it is observed that the changes in the image Laplacian along the direction perpendicular to the gradient are different in the inpainted and untouched areas.In order to better show the difference between the two,the weighted least square filter enhancement algorithm is introduced to filter and enhance each color channel of the inpainted image,which captures the impact of image inpainting from multiple perspectives.On this basis,a feature set is constructed based on the intra-channel and inter-channel local variances of the changes to identify theinpainted areas,and then it is sent to an ensemble classifier for classification decision.Finally,the results of inpainting location are further improved by abnormal exposed regions exclusion and morphological filtering.Through a large number of experimental results,it can be found that when the inpainted area is small,the detection performance of this algorithm is better than the existing methods,and it has better robustness to common post-processing(such as gamma correction,scaling,etc.).(2)Considering the inherent differences between the inpainted and the untouched areas in texture,and the high time complexity of the weighted least square filtering,and in order to retain more small details of texture,this paper further proposes to use gradient domain guided image filtering to enhance the inpainted image,which can better retain the texture structure of the untouched areas,and also highlight to a certain extent The fuzzy effect of the inpainted areas.In addition,compared with the weighted least square filtering algorithm,the complexity of the gradient domain guided image filtering algorithm is relatively low.The experimental results show that the algorithm based on the gradient domain guided image filter enhancement has better detection performance for the inpainting forgery areas,and the running time is about 6 times shorter than the weighted least square filter algorithm.Compared with the existing method,the algorithm can still detect and locate the inpainted areas better after the post-processing such as gamma correction and scaling.