Use this URL to cite or link to this record in EThOS:
Title: Learning based forensic techniques for source camera identification
Author: Li, Ruizhe
ISNI:       0000 0004 6061 9470
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Access from Institution:
In recent years, multimedia forensics has received rapidly growing attention. One challenging problem of multimedia forensics is source camera identification, the goal of which is to identify the source of a multimedia object, such as digital image and video. Sensor pattern noises, produced by imaging sensors, have been proved to be an effective way for source camera identification. Precisely speaking, the conventional SPN-based source camera identification.has two application models: verification and identification. In the past decade, significant progress has been achieved in the tasks of SPN-based source camera verification and identification. However, there are still many cases requiring solutions beyond the capabilities of the current methods. In this thesis, we considered and addressed two commonly seen but less studied problems. The first problem is the source camera verification with reference SPNs corrupted by scene details. The most significant limitation of using SPN for source camera that SPN can be seriously contaminated by scene details. Most existing methods consider the contaminations from scene details only occur in query images but not in reference images. To address this issue, we propose a measurement based on the combination of local image entropy and brightness so as to evaluate the quality of SPN contained by different image blocks. Based on this measurement, a context adaptive reference SPN estimator is proposed to address the problem that reference images are contaminated by scene details. The second problem that we considered relates to the high computational complexity of using SPN in source camera identification., which is caused by the high dimensionality of SPN. In order to improve identification.efficiency without degrading accuracy, we propose an effective feature extraction algorithm based on the concept of PCA denoising to extract a small set of components from the original noise residual, which tends to carry most of the information of the true SPN signal. To further improve the performance of this framework, two enhancement methods are introduced. The first enhancement method is proposed to take the advantage of the label information of the reference images so as to better separate different classes and further reduce the dimensionality. Secondly, we propose an extension based on Candid Covariance-free Incremental PCA to incrementally update the feature extractor according to the received images so that there is no need to re-conduct training every time when a new image is added to the database. Moreover, an ensemble method based on the random subspace method and majority voting is proposed in the context of source camera tackle the performance degradation of PCA-based feature extraction method due to the corruption by unwanted interferences in the training set. The proposed algorithms are evaluated on the challenging Dresden image database and experimental results confirmed their effectiveness.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: QA76 Electronic computers. Computer science. Computer software ; T Technology (General)