Research On Identification Of Digital Image Source Equipment

With the rapid development of multimedia technology,the number of image is fast growing.Digital images have brought joy and convenience to people's life.However,it also bred a series of social problems and potential security issues.For example,the content of digital images can be tampered easily which makes people begin to question the reliability of the image source and the authenticity of image content.To this purpose,digital image forensics technology is coming.Digital image source identification technology is a branch of digital image forensics technology which aims at tracing the origin of image by exploiting intrinsic artifacts left on the acquired image at shooting time by the acquisition process.Digital image source identification is considered as a multi-classification problem generally.Traditional algorithms mainly classify the camera model by handcraft features.In this paper,two algorithms for digital image source identification are proposed.The algorithm based on multimodal information and convolutional neural network and the algorithm based on feature fusion.For these two algorithms,the following research results are obtained:(1)The algorithm based on multimodal information and convolution neural network is proposed.In this algorithm,we add two modes of original image by pre-processing operation of Gauss filtering and extracting image pattern noise.Then we use these three modes of image as the input of a three channel convolutional neural network separately.Then three independent models are trained in their respective channel.At last,we use the combinatorial output as the final output.We also optimized each single channel network by adding the multi-level feature extraction structure on the basis of traditional network.Finally,the effectiveness of the proposed algorithm is verified by experiments.(2)The algorithm based on feature fusion is proposed aiming at the problem that a single feature cannot completely depict an image.We explored how the impact is on classification performance if we utilize the multisource of features extracted by CNN and tradition handcraft feature.We choose the image pattern noise feature and co-occurrence matrix feature as the handcraft feature in this paper and extracted the CNN feature from the model mentioned above.Feature level fusion and decision level fusion are then used to combining these two features.We use the support vector machine(SVM)to do the classification work.Finally,we present experimental results showing that our framework can identify the correct model of an image's source camera with a higher accuracy than the algorithm using single feature.(3)Furthermore,we use the CNN as the feature extractor and then SVM is used to do the next classification work.The experiment has shown that classification results is better than the simple CNN.