Camera Source Identification Based On Limited Labeled Training Set

With technological development, as well as the popularity of various types of digital imaging equipment, photos have become an important way to send information. But with the rise and widely used of a variety of image editing software, in addition to the increasingly open and inclusive network environment, tampering digital images are increasing. The digital images we see may go through the secondary compression, stitching, copying and other operations. Seeing no longer is believing. In recent years more and more events about digital image forgery have occurred, these events raises a concern to digital image authenticity and originality. At the same time, these also makes digital image forensics research in recent years become a hot topic.Existing camera sources forensics methods don't work well under the conditions of insufficient prior knowledge. This paper focus on this problem and propose methods to solve this problem. Since the source forensics is divided into camera model identification and camera individual identification, and the features they using are not same. So in this paper, the solution is given to solve the insufficient of training samples. The the main work and the main results are as follows:(1) Proposed a camera model source identification method based on ensemble projectionIn the existing digital camera image source model identification methods, local binary pattern feature works well. But, when the number of training samples is insufficient, the identification accuracy has a significant decline. So, in this paper, ensemble projection method is given to solve this problem. First, training the S VM using random dimension of local binary patterns features with limited labeled training samples. Then using the SVM to classify all samples, and higher posterior probability samples were selected to build a prototype set. This process is repeated several times to obtain several prototype sets which represent a part categories information; Secondly, the label samples were mapped on the prototype set and the value of posterior probability is regarded as projection value. And then all projection values were connected to obtain the final projection features. Then, using labeled samples'projection features to train classifier and using this classifier to classify unlabeled samples. Finally, experimental results illustrate that the ensemble projection method achieves a notable higher average accuracy than previous algorithms when labeled training samples is limited.(2) Proposed patch-based sensor pattern noise for camera individual source identificationIn the existing digital camera image individual source identification methods, the most common feature is the sensor pattern noise (SPN). Typically, for each camera, the number of training samples is no less than 50. When the number of training samples is insufficient, the identification accuracy rate will be seriously low. So, in this paper, we propose patch-based sensor pattern noise to solve this question. Firstly, the training images are divided into blocks according to a certain size, and for each location the blocks with lower complexity coefficient are selected to get local reference pattern noise. Then combining all local reference pattern noise blocks and getting a full reference pattern noise. In the test phase, the image block with lowest complexity coefficient are selected to make correlation with corresponding block in the reference pattern noise, and determine the source of the camera by the threshold. Finally, experimental results illustrate that the proposed patch-based method has well performance, when labeled training samples is limited...