Digital Camera Source Identification Based On Local Pattern Noise And Small Training Set

The continuously emerging of digital image forgery events subvert the traditionalconcept of seeing is believing and cause negative effects on political, news, scientificresearches and so on. So it urgently needs to determine the authenticity and integrity of digitalimages. Digital image forensic technology, which is of very important significance to ensurepublic trust order and fight against crime, emerges as the situation requires. As an importantbranch of digital image forensic, digital image source identification aims to determine thesource of digital images.Based on the simulation and analysis of typical digital image source identificationalgorithms, this paper mainly research digital camera source identification based on correctiondetection of pattern noise and pattern classification. The main purpose is to improve theaccuracy and reduce the quantity of training set of image source identification algorithm.For the digital image source identification based on pattern noise, an algorithm of digitalimage source identification based on pattern noise from local areas was proposed. Throughthe analysis of the influence caused by image content, we found that pattern noise from areaswith high brightness and low texture has a better quality. In order to overcome someshortcomings such as complex operations and needing additional training set appeared in thesimilar algorithm, an evaluation function was proposed to select the areas with highbrightness and low texture, and pattern noise from these areas were used for digital imagesource identification. Experimental results have demonstrated the effectiveness of theproposed algorithm.For the digital image source identification based on pattern classification, most methodsneed large training set, which may not be satisfied in reality. So an algorithm for digital imagesource identification based on small training set was proposed. The training set was firstsegmented into overlapped blocks to enlarge sample space, and then effective features wereextracted from each block, finally a SVM classifier was trained, which can be used for digital image classification. The proposed method is still effective when there are very small trainingset. The conclusion of wavelet features that extracted from training set with high texture aregood for improving the performance of the proposed method was also demonstrated.