Research On Digital Image Resampling Operation And Splicing Operation Forensics Based On Deep Learning

With the rapid development of Internet technology,digital image processing technology is also becoming more and more developed.Nowadays,people can easily edit digital images using a variety of software in their lives.Although these modified images can bring us some visual or entertainment effects,there will still be some malicious tampering,and these malicious tampering images will have a huge impact on litigation evidence collection,criminal investigation,national political and military,etc.Therefore,it is of great significance to studying the image forensics task.However,tampered images are usually edited in various ways.In this paper,we mainly focus on studying two aspects of whether resampling operation and digital image splicing operation.The main works are as follows:At present,it is difficult to detect the resampling operation in the recompressed image,especially in the down-sampling operation,this paper analyzes and designs a three-stream convolutional neural network based on the existing detection resampling algorithms,which is used to improve the accuracy of detecting resampling operations in the recompressed images.First,an image tampering detection method based on a three-stream convolutional neural network is designed.Then,a custom high-pass filter and a restricted convolution kernel are used to learn the resampling features on the recompressed image and suppress the recognition of recompressed image content and stack the features extracted from the three streams.Finally,considering that the features of the lower level largely reflect the traces of the resampling operation,the features of the lower level are fully retained by the skip connection,which greatly improves the final detection accuracy.Experiments are compared with current popular methods.In the case of testing the same data set,the proposed method can effectively detect the resampling operation in the recompressed image,especially the accuracy of detecting the down-sampling operation is improved by 5%-7%.Experimental comparisons have also been made on the overall localized resampling region.The results demonstrate that the algorithm proposed in this paper has surpassed the existing method on the down-sampling recompressed image,which makes it more accurate to locate the tamper area in the whole image.The image splicing operation is a very common tampering method,but the existing methods can not accurately detect and locate the splicing area.Therefore,we design a digital image splicing detection algorithm focusing on the splicing area and the edge information of the area under the guidance of inconsistency.This algorithm consists of three branches.First of all,in the process of tampering,the edge of the splicing object will leave traces of operation.In order to enable the network to learn the edge information of the image,it is difficult to make the network converge because the pixels of tampering object edge is too few.Therefore,this paper expands 6 pixels inward and outward along the edge of the spliced object respectively,and blackens it to form a doughnut,which is used to learn the inconsistency of the edges of the splicing area.Secondly,the information contained in each image is different due to the light noise of the camera equipment at the time of image capturing,etc.,which leads to the difference in the hidden information from different areas of the image when the splicing object is copied from the original image to another image.Therefore,this paper designs a branch to focus on the difference between the splicing area and other areas.The branch introduces the attention mechanism for the first time to focus on the learning splicing area.Finally,a binary output branch is designed,where 0 represents the untampered images and 1 represents the splicing images.The branch can quickly and effectively detect whether the input image is a tamper image,and can be combined with the results obtained by the above two branches to determine the final detection results.The experimental results show that the network model designed in this paper not only has a good generalization ability for different resolution images,but also has better generalization performance for the tampering images copied and pasted on the same image.The method in this paper is superior to the existing methods in detection accuracy and position precision.