Study On Machine Learning Based Algorithm For Distinguishing Natural And Computer Generated Images In Digital Forensics

With the rapid development of digital media technology,the capacity of using complex mathematical models to produce high realistic computer synthetic images is increasing day by day,which greatly promotes the development of virtual reality technology.However,it also brings security problems such as the authenticity authentication of digital images.How to discriminate natural images and computer-generated images is one of the important topics in digital image authenticity identification.Most of the existing methods use sensor pattern noise as the image forensics feature,but its descriptive ability is insufficient.This dissertation combines dual-tree complex wavelet(DT-CWT),support vector machine(SVM)and deep convolutional neural network(DCNN)in machine learning,and proposes a set of solutions.The specific research work is as follows:1.A natural image and computer generated image discrimination algorithm based on DT-CWT domain and SVM is proposed.In this algorithm,the denoising operator is used to obtain the sensor pattern noise of the image,then singular value decomposition(SVD)is utilized for enhancement.Next,the energy and deviation of each sub-band are extracted as the characteristics based on the DT-CWT.Finally,SVM is used for classification.Experimental results show that this method not only has lower feature dimension,but also has higher accuracy.2.A natural image and computer-generated image identification algorithm based on DCNN is proposed.In the pre-processing stage,the denoising operator in algorithm 1 is used to generate the pattern noise image.In the feature extraction step,the 8-layer DCNN designed in this paper is used to extract the features of the pattern noise images.Finally,SVM and SoftMax are used as classifiers.Experimental results show that compared with existing literatures,this algorithm has the advantages of simple network structure,fast training speed and can achieves high accuracy.3.This paper proposed a natural image and computer generated image recognition algorithm based on transfer learning.In this algorithm,the denoising operator in 1 is firstly used to generate the pattern noise image,and then the pre-training models like AlexNet,VGG16 and ResNetl8 are transplanted and retrained on the small-scale image set by means of fine-tuning.Through experimental verification,the problem of image recognition in this paper is solved by fine-tuning parameters of the trained deep neural network,which can not only meet the requirements of small sample data,but can also achieve higher accuracy.