Digital Image Forensic Techniques For Source Camera Identification

Digital image forensic technology is used for identifying digital image tampering, forgery and stego. It has great political significance to carry out this research work to ensure public confidence, combat crime, safeguard judicial justice and information integrity. Digital camera source identification problem is an important research direction of digital image forensics technology. It analyzes and judges the machinery within the digital image which arises from the imaging process of imaging device. Currently, the there are two ways to trace the source device: First, using equipment fingerprint to implement correlation detection; Second, pattern classification based on image characteristics. Two ways have their own advantages and disadvantages, there are certain differences on the applicable applications.In this paper, we carry out the research work through correlation detection and pattern classification. We mainly focus on how to improve the recognition rate of correlation detection , how to reduce the computational load and how to further improve the performance of pattern classification algorithm .We simulate some classical algorithms and analyze the advantage and the disadvantage of these algorithms firstly. Then based on this, we propose two innovated algorithms:1. For the correlation detection algorithm, during the extraction and pre-processing of pattern noise, we discuss how the de-noising filter and the de-CFA interpolation operation can improve the quality of pattern noise, we compare the performance of four filters. Afterward, we focus on discussing why and how the use of large components of the sensor pattern noise can improve the performance of the correlation-based detector. The proposed method not only further improve the recognition rate of the correlation-based detector and the robustness to JPEG compression, but also greatly decrease the computational load for detection.2. For the pattern classification algorithm, we first analyze the independent classification ability of each category feature which extracted by existing algorithm, select the category from which some of better performance characteristics, and then apply the major component of pattern noise extraction method to improve the features based on the difference statistics. Combined with features extracted from scanner classification based on pattern recognition algorithm, we increase the number of wavelet features to further improve the classification performance. This algorithm improves the recognition rate of pattern classification algorithms.