Research On Image Splicing Tamper Detection Method Based On Deep Learning

With the development of digital media technology,digital images are becoming an indispensable part of people's daily work and life,while deriving many powerful image processing software,so that people can easily complete the image modification.However,while satisfying people's aesthetic needs,the development of this kind of software also brings some disadvantages that cannot be ignored.The illegal use of tampered images has caused serious damage to the credibility of public utilities,especially tampering with real images in military,political and other related fields may even lead to unnecessary military conflicts and political disturbances,the appearance of these false images has led to many doubts about the authenticity and information integrity of the current image content.Therefore,it is very important to develop an effective identification method for the authenticity of images.Forensic research on digital image tampering is a hot topic in the field of information security in recent years,as one of the most common forms of digital image tampering and forgery,image splicing tamper detection is an important content of image tampering forensics research.The research shows that the traditional image splicing tamper detection method is difficult to achieve ideal results due to the multiple influences of light,noise and other attributes.In view of the remarkable innovative achievements of deep learning in the field of computer vision,this paper studies how to use deep learning to achieve image splicing tamper detection and location based on the existing image splicing tamper detection technology.The main tasks of this paper are as follows:1.In order to effectively improve the accuracy of image splicing tamper detection and simultaneously locate the tamper area of image splicing,an image splicing tamper detection model based on attention mechanism was designed.The model is mainly composed of two parts.One is to improve the convolutional neural network combined with the attention mechanism to help the model focus on some local information of the image and extract effective tamper features,so as to achieve better detection effect.The other is to locate the tampered areas in the image by using the adversative complementary learning(Acol)method and the Gradient-weighted Class Activation Mapping(grad-CAM)method,and then map them to the tampered map in the form of a thermal map to enhance the interpretability of the location results.Through the comprehensive analysis and comparison of experimental data,it is shown that the model has high accuracy and generalization ability.2.In order to further improve the tampering region localization effect of image splicing tamper detection,a tampering detection and localization model based on two-channel dilated convolution was proposed.In this model,the noise channel is mainly used to extract various image noise features through a set of high-pass filters,and then the real image region and the tampered region are distinguished according to the consistency of noise features.RGB channel combines dilated convolution with attention mechanism to extract RBG image features and better locate tampered areas.Then,the features extracted from the two channels are fused through bilinear pooling.Finally,the fusion features are used to complete tamper classification and boundary box regression,so as to realize tamper region location.The experimental results show that this method has better localization performance than the existing tampering region localization algorithm.