Feature Set Reduction for Source Camera Identification Using Demosaicing and Wavelet-based Features

Recent research work done in Camera Forensic Techniques have exhibited excellent results by merging several modern methods of Source Camera Identification, thereby creating a large feature set. Increasing the feature set can also contribute to higher computational time, complexity and error rates due to use of redundant features. The work exhibited in this paper shows that it is possible to optimize the feature set size to 16, while maintaining the results as those obtained by state-of-the-art techniques. The statistical features extracted from images are based on demosaicing algorithm and wavelet transform. First, the proposed statistical equations were used to formulate 4 features based on periodic characteristics shown by demosaicing. Then, the image matrices were transformed by 1-Level wavelet decomposition and 16 features were extracted from obtained co-occurrence matrices. After further optimization, it was found that classification model trained using 16 features had the highest average accuracy of 99.2%. Thus, the final optimized feature set size was reduced to 16 features from 20 features for identifying images from 10 camera models.