Research And Application Of Source Camera Identification Algorithm Based On Pattern Noise

With the continuous development of image acquisition equipment and innovation of image processing technology,images have gradually become one of the main medias of daily communication.It brings great convenience for obtaining and editing of images.Meanwhile,there are many latent danger of information security.To achieve their illegal purposes,some lawbreakers wantonly tamper and forge the digital images involved in news reports,medical images and even in court evidence.This affects the public order and customs of the society seriously,which reduces the credibility of the government.In this context,source camera identification(SCI),which aims to identify the source and authenticity of images,is of great significance.Among different technologies in the field of digital image forensics,this dissertation focuses on the SCI task at camera instance level.Its purpose is to tell the origin of camera instance that shoot the current investigated image,which is a helpful way for solving the problems of image traceability and copyright protection.Based on analysis of existing problems in the overall source camera identification framework,this paper carried out work on pattern noise extraction,residual aggregation in traditional signal processing methods.Guided by prior knowledge that spatial information is helpful in SCI problems,effective deep learning network design for instance level feature representation is also studied.The main work of this dissertation includes:(1)A source camera pattern noise estimation algorithm based on multi-scale expected patch log likelihood denoising is proposed in this dissertation.Firstly,the multi-scale expected patch log likelihood(MSEPLL)denoising algorithm is adopted to obtain accurate pattern noise,as well as to reduce the impact of image content on the residual image.Secondly,the visual effect,identification accuracy and receiver operating characteristic curve(ROC curve)are used to measure the effectiveness of the algorithm based on the proposed device level source camera identification framework.Finally,a large number of experiments are carried out to verify the effectiveness of the proposed algorithm.Performance improvement for small image patches are more prominent,which is crucial for real applications.(2)A camera pattern noise aggregation algorithm based on edge guided weighted average and maximum likelihood estimation is proposed in this dissertation.Firstly,the edge guided weighted average method is adopted to set higher weight coefficients to pixels in more reliable regions in the estimated residual image.Furthermore,a weighted normalized correlation coefficient is proposed to measure the similarity between the camera fingerprint and the test image fingerprint in the test stage,which improves the overall identification accuracy.On this basis,the maximum likelihood estimation method is integrated to the residual aggregation process to improve the accuracy of camera fingerprint estimation.Then,the effectiveness of the proposed edge guided weighted average method and the later maximum likelihood estimation fused version are compared respectively.Finally,A source camera identification software based on pattern noise is presented in this dissertation.The above source camera pattern noise estimation algorithm and camera pattern noise aggregation algorithm are integrated in the software.Functionalities of camera fingerprints estimation,fingerprints estimation from single test image,similarity calculation and source camera identification are implemented to facilitate real forensic applications.(3)A source camera identification algorithm based on residual prediction module and convolutional neural networks is proposed in this dissertation.Firstly,a residual prediction module is adopted to reduce the influence of the image content,so as to improve the robustness of camera instance features.Secondly,under the guidance of the prior knowledge from conventional signal processing methods,pooling layers are removed and a convolutional neural networks module is proposed to retain the spatial relationship of image features.Finally,large-scale experiments are carried out to verify the effectiveness of the proposed convolutional neural network module.