Digital Image Region Duplication Forgery Detection Algorithm Based On Deep Learning

Digital image blind forensics is an important research direction in the field of multimedia security in recent years.Traditional image blind forensics technology usually uses artificial extraction of the statistical features of forgery traces to detect and loclize the forged area of the image.However,the operation of forged image is more complex and hidden,and the imperceptible trace features are often destroyed by various post-processing operations,which brings great difficulties and serious problems to forensic work.With the development of artificial intelligence technology,blind image forensics using machine learning has become a trend.In this paper,deep learning technology is used to study the blind image forensics method for region duplication forgery.The main work of this paper is as follows:(1)This paper analyzes and summarizes the existing image region duplication forgery detection algorithms,models the image region duplication forgery,and describes the general image region duplication forgery detection process.This paper expounds the influence of convolution and pooling on the field of vision of convolution neural network,analyzes the problems caused by different sizes of tampered areas detected by the algorithm and puts forward solutions.Expand the existing region duplication forgery datasets.(2)Aiming at the problems of single field of view and insufficient field of view,a two-branch network-based image area duplication and forgery detection algorithm is proposed.The detection network consists of preprocessing module,branch network module and fusion module.Firstly,the image is preprocessed and the image size is normalized.The branch network module uses two branch networks to localize the tampered area of the image.The feature extractors of the two branch networks have different fields of vision.Then the distance measure is used to sharp the tampered traces of the image.Finally,the decoder is used to localize the tampered area of the image.The fusion module concatenate the results of the two branch network,and uses the classifier to classify the image pixels to localize the tampered area.The experimental results show that the algorithm can not only resist Gaussian noise,geometric attacks and JPEG compression,but also has certain effectiveness for different size of the copy area.The algorithm can achieve 56.64%precision and 61.24%recall on the CASIA V2.0 copy-move tampering dataset.The average detection time is 1.36 seconds.(3)Aiming at the problem of slow detection speed and low detection accuracy of existing methods,a two-dimensional CNN-based image region duplication forgery detection algorithm is proposed.The algorithm uses two-dimensional CNN to extract image features,measures the distance of the extracted features,and sends the result of distance measurement to the classifier to determine the authenticity of the image.Experimental results show that the algorithm can distinguish whether the image is a region duplication forgery image end-to-end,the detection speed is fast,the detection accuracy is high,and it has a certain resistance to geometric attacks.The algorithm can achieve 98.5%accuracy on MICC-F 220 dataset.(4)Aiming at the problem that the existing methods perform well in the training dataset and poorly in the test dataset,an image region duplication forgery detection algorithm based on 1D CNN is proposed.The algorithm combines the accuracy of traditional methods to extract features and the accuracy of deep learning method to complete the image region duplication forgery detection.The algorithm uses two branch network to distinguish the authenticity of the image.The two branch networks obtains the image features through preprocessing,sharps the image tampering traces through distance measurement,and finally sends them to one-dimensional neural network to extract the image distance measurement features to distinguish the authenticity of the image.The experimental results show that the detection result of the fusion network is better than that of any single branch network,the algorithm has strong resistance to geometric attacks,and the detection result is stable and reliable.The algorithm can achieve 97.0%accuracy on the MICC-F 220 dataset.