Digital Image Source Forensics Method Based On Multiple LBPs In Multicolor Spaces And Convolutional Neural Network

With the rapid development of computer graphics processing technogy,digital image source forensics has received increasing attentions.An enormous variety of different schemes have been proposed for digital image forensics,especially for identifying natural images and computer graphics.As a normal feature for identification,texture feature is an image-based,easy to calculate and the most representative feature of images.However,there are still many deficiencies in the published schemes,many of those schemes lack of capacity of global features for image description.Moreover,because the outstanding performance in image recognition,convolutional neural network(CNN)has been applied on image forensics in recent years.The schemes based on CNN are easy to calculate and can learn features automatically,which are better than typical schemes.In this thesis,some existing schemes for identifying natural images and computer graphics have been reviewed and two novel methods are proposed.The main contributions of this thesis are as follows:1)A review of the research works about typical image forensics schemes based feature extraction and classification and the deep learning methods for image identification is given.Those schemes based on texture feature and convolutional neural netwok are introduced.The methodology of these schemes are useful for the future works.2)A digital image forensics scheme based on multiple local binary patterns(LBP)in multicolor spaces is proposed for distinguishing natural images from computer generated graphics.The proposed scheme extracts texture features from micro and macro level of image and features from multicolor spaces for identification.The simulations gives an ideal performance and shows the proposed method is robust against JPEG compression,resizing,rotation and adding noise.3)A digital image forensics scheme based on CNN and transfer learning is proposed.The CNN-based method can learn and extract features automatically.In addition,transfer learning makes this scheme more flexible.The proposed schemes in this thesis possess a high correct detection rate of CG,and they are robust against some post-treatments on the image.Both these two schemes provide us a novel way for image forensics.