Image Forgery Detection Based On Deep Convolutional Neural Network

As an important means of information transmission,digital image plays an important role in social development.With the rapid development of computer technology and digital media technology,the emergence of a large number of image editing software makes it easy for people to tamper with the content information of the image,which makes the credibility of image drop sharply.These frequently appearing tampered images have caused serious adverse consequences in the military,judicature,media and other fields,and have also caused researchers to pay attention to image tampering.At present,the detection algorithms for image splicing forgery are mainly image forgery detection algorithm based on feature extraction.The feature extraction based image forgery detection algorithm utilizes a certain attribute of the image(eg based on image hash attribute and based on imaging device attribute,etc.),and the corresponding drawback is that the detection algorithm,when the attribute is destroyed or the information strength is insufficient,tends to be in a failure state.In this thesis,based on the above mentioned problems,the image feature extraction based image forgery detection algorithms are deeply studied.The main works include the following aspects:1.Based on the characteristics and design ideas of convolutional neural networks,we further design and propose a cascade network structure of shallow sparse neurons to replace the single network structure of deep multi-neurons(SNDM),and make up for the drawbacks of SNDM by re-learning the characteristics of difficult samples.The detection algorithm consists of two parts: cascaded convolutional neural network and adaptive filtering post-processing.Experimental results show that the detection algorithm has better detection effect and higher robustness.2.Firstly,the residual propagation theory in the residual network is studied,and then the residual feedback module structure is further designed according to the above idea of residual propagation,so as to achieve the purpose of amplifying the hidden feature differences between the tampered and un-tampered regions,and suppressing those characteristics which may be misjudged.The proposed ringed residual network(RRU-Net)adds the residual propogation module structure on the basis of the classical U-shaped network(U-Net)to combat the gradient disappearance problem in the image forgery detection task(the gradient disappearance refers to the disappearance of the gradient of the hidden feature differences between the tampered and the un-tampered regions),and add the residual feedback module structure to enhance these hidden feature differences.Experimental results show that the image forgery detection algorithem has better detection effect and higher robustness.