Research On Source Camera Identification Based On Expert Prior Knowledge

With the development of multimedia technology,digital images have gradually become one of the main ways for people to receive and disseminate information.At the same time,the emergence of various powerful image editing software has caused people to worry about the integrity and authenticity of digital images.In this case,source camera identification,which aims at identifying the source camera of an image,has attracted a wide range of attention.Most existing methods only focused on extracting features from the single artifact of the camera left on the captured images and ignored other artifacts that may help improve final accuracy.In this paper,we propose a general framework for source camera identification.Based on this framework,three source camera identification methods are proposed:(1)Proposed two methods for source camera identification using fusion featuresFirstly,a source camera identification method based on ensemble classifier is proposed.This method combines three different features,which are extracted from the sensor pattern noise,the estimated parameters of CFA interpolation algorithm and the extracted features from the original image.These three features can comprehensively capture the various artifacts left by the camera on the image.Then the problem of slow training of high-dimensional features is solved by using an ensemble classifier.Secondly,we proposed a source camera identification method based on rich model.This method avoids capturing camera software-related artifacts by estimating camera software algorithms and their parameters,and by combining multiple submodels in the pre-processing part to build a rich model to capture camera legacy information.Experimental results prove that the performance of the two proposed methods are better than the existing feature-based source camera identification methods.(2)Proposed a source camera identification method based on deep learning combined with domain experts' prioriThis paper introduces deep learning into the proposed general framework,and uses convolutional neural networks to complete the tasks of image transformation,feature extraction and classification in the proposed framework,avoiding the use of complex image transformation algorithms and feature design.Before sending the image to the convolutional neural network,we first use a preprocessing module composed of denoising algorithm,demosaicing algorithm and predicting algorithm to process the image.The preprocessing module can introduce a priori knowledge of domain experts to capture camera hardware and software related artifacts,thereby suppressing the impact of image content on source camera identification and effectively avoiding the loss of camera attribute information.The experimental results show that the proposed deep learning-based method can effectively identify the source camera brand,model and individual of the image,and its overall performance is better than the two methods using manual features and the method directly using convolution neural networks without introducing a priori knowledge of domain experts.