Source Camera Identification Based On Optimal Utilization Of Limited Samples

Nowadays,there are various advanced products and technologies in the society,such as high-performance smartphones,sophisticated image editing applications and widely used social media applications by the public.Their appearance makes it easy to capture,edit or propagate images for amateurs.And they make the network flooded by all kinds of digital images.The development of high technology as a double-edged sword,while enriching and convenient life,also makes the credibility of digital images gradually reduce,there will be criminals to achieve their own purposes and malicious tampering with images.In order to further ensure the credibility of digital image,the passive digital image forensics technology proposed by researchers plays an irreplaceable role in forensics.As a significant branch,the Source Camera Identification(SCI)aims to identify the origin device of given digital images.The Source Camera Identification usually trains the classifier through the labeled image samples,and then identifies the unlabeled image source.However,in practical application,it is a big challenge to find enough labeled image samples.In the case of limited labeled image samples,most of the existing methods will lose the original high accuracy and are not reliable.In order to solve this problem,this paper proposes two specific solutions,the main research contents are as follows:(1)Firstly,a Source Camera Identification method of limited sample based on distance semi-supervised ensemble learning is proposed.Based on the European distance,Manhattan distance and Chebyshev distance between the two samples,different thresholds are set to select the unlabeled image samples near the labeled image samples.Then,through ensemble learning,the unlabeled image samples with high frequency of occurrence near the labeled image samples are selected to stack the labeled image samples to form a new training set.Then the pseudo label is self-corrected by SVM to improve the authenticity of pseudo label of unlabeled image samples,which makes the new training set image sample label more pure,and then achieves higher classification accuracy.The experiment shows that the Source Camera Identification method of limited sample based on distance semi-supervised ensemble learning can make full use of the unlabeled image sample information,and can greatly improve the Source Camera Identification accuracy in the case of limited samples compared with other methods.(2)A Source Camera Identification method of limited sample based on ensemble learning projection is proposed.LBP and CFA features are introduced to further enrich the prior information when given the labeled image samples.And constructing multiple feature subsets to establish multiple pre-classifiers.Then the semi-supervised method is introduced and all image samples are tested by pre-classifiers.The posterior probability is taken as the measurement standard,and the image samples with higher posteriori probability are selected to form the prototype set.Through the training of the classifier,the posterior probability of each image sample belonging of each class are obtained,which is taken as the final projection vector,also is the new classification feature.Finally,many weak classifiers are trained by using new features,and the final Source Camera Identification results are obtained by the ensemble learning voting method.The experimental results show that this method has good performance and can improve the Source Camera Identification accuracy in the case of limited samples.