Research Of Camera Model Identification Algorithm Based On Convolutional Neural Network

The advances in image acquisition equipment and network technology,as well as the development of image processing technologies,make images gradually become an important carrier of communication.However,it also provides an opportunity for criminals.If tampered images are applied in news reports,court evidence,insurance claims,international affairs,etc.,it will cause serious social problems and conflicts.Digital image forensics technology,which aims to use various image features to analyze images to verify the authenticity of images,is receiving more and more attention.As a convenient and direct image forensics technology,camera model identification can determine the camera model of the image to be checked through the hidden camera features in the image without adding additional information.This method has become an important research subject in image forensics.In addition,the accurate camera model identification algorithm can effectively assist manipulation detection and others,it has great application value in the field of forensics.This paper focuses on the problems in the current research,such as the camera model features are easily affected by image content,and the performance of various algorithms is difficult to compare.Based on convolutional neural networks,different preprocessing modules are designed.Under fair and reasonable experimental setup and performance evaluation standards,the camera model identification is studied.This paper mainly introduces the works from the following aspects:(1)A camera model identification algorithm based on multi-feature representation and richer convolutional features network was proposed in this paper.First,the multi-feature representation consists of a multi-scale high-pass preprocessing module and RGB color channel information.The multi-scale high-pass preprocessing module can enlarge the camera model features.On this basis,RGB color channel information is fused to provide more interpolation information and enrich the camera model features.Secondly,the modified richer convolutional features network makes full use of the features learned in the network to obtain rich features related to the camera model.Ultimately,it improves the effect of camera model identification based on image patches.(2)A camera model identification algorithm for large scale data sets was proposed.The network of this algorithm has a small number of parameters,but a satisfactory identification effect can be obtained.First of all,the ResNet-based residual extraction module is used to enlarge the camera model features in the original images.Furthermore,the modified SqueezeNet network is used to extract camera model features.Finally,the experiments at the image patch level prove that this algorithm can significantly improve the identification ability of camera models.(3)A source camera identification algorithm based on the selection of diverse image patches and the residual prediction module was proposed.This algorithm can perform camera model identification,and it is also effective for camera brand or camera instance identification.In the first place,this algorithm selects representative training and validation data from each training and validation image based on multiple image patch selection criteria.It can ensure that this algorithm can extract richer camera features from limited data to obtain better identification result.Secondly,in order to reduce image content interference,a residual prediction module based on fine-grained multi-scale characteristics is designed,and then the modified VGGNet-16 network is used for camera feature extraction.Eventually,this algorithm can obtain better results at three different levels of camera identification tasks.