Detection and localization of forgeries in digital images

Simple item page
dc.contributor.authorAhmed, Belal
dc.contributor.supervisorGulliver, T. Aaron
dc.date.accessioned2021-01-05T20:46:18Z
dc.date.available2021-01-05T20:46:18Z
dc.date.copyright2020en_US
dc.date.issued2021-01-05
dc.degree.departmentDepartment of Electrical and Computer Engineeringen_US
dc.degree.levelDoctor of Philosophy Ph.D.en_US
dc.description.abstractDigital images have become a dominant source of information and means of communication in our society. However, these images can easily be altered using readily available image editing tools. Image tampering can be done in several ways such as image splicing, retouching, and copy-move forgeries. In copy-move forgery, part of an image is copied and pasted into a different part of the same image for the purpose of hiding or adding an object to the image. In image splicing, part of an image is copied and pasted into a different image. To detect image forgeries, image features must first be extracted. A feature is information related to the edges, objects or a specific region in the image. In this dissertation, new methods for detecting copy-move forgery and image splicing are introduced. Most existing block-based forgery detection methods use large feature dimensions up to 64 per image block so the complexity is high. However, reducing the feature dimensions lowers the detection accuracy, so a new method of detecting copy- move forgery in images using only 4 features per image block. This method uses steerable pyramid and singular value decomposition (SVD) techniques to decompose and extract features from image blocks. Then the features are sorted lexicographically and matched using the Kolmogorov-Smirnov (KS) test. The proposed algorithm is compared to several well-known techniques and shown to provide better accuracy. To detect image splicing, a new deep learning method is introduced. This method employs Mask-RCNN to generate masks for spliced regions in forged images. It is specifically designed to learn discriminative artifacts from tampered regions. In this method, a ResNet backbone is used to convert the input image into a feature map. The ResNet-50 and ResNet-101 backbones are considered. The ImageNet, He_normal, and Xavier_normal initialization techniques are employed and compared based on convergence. To train a robust model, several post-processing techniques are applied to the input images. Several techniques have been introduced for image forgery detection. However, most only focus on detecting a certain kind of forgery and perform poorly in other cases. As a result, detecting multiple kinds of forgery using one technique remains a problem. Thus, a novel deep neural architecture called PADNET is introduced which has been specifically designed to detect multiple kinds of forgery. Unlike other solutions, PADNET is an end-to-end trainable deep neural network which employs feature pyramid network (FPN) to aggregate features from multiscale levels of a ResNet-50 backbone. The feature maps are then used to train a DeepUNet architecture designed to learn discriminative features by considering both high-level global features and low-level local features. The convergence of PADNET is tested using two loss functions, binary cross-entropy and weighted binary cross-entropy. Experimental results show that weighted binary cross-entropy is more efficient as a loss function for copy-move forgery while binary cross-entropy is more efficient for image splicing. In addition, the performance of PADNET with training on only the boundaries of the forged area is compared to the network trained on the entire forged area. Evaluation is done using the well known CoMoFoD dataset for copy-move forgery and CASIA1 for image splicing forgery. The results obtained demonstrate that PADNET outperforms state-of-the-art copy-move and image splicing forgery detection algorithms.en_US
dc.description.scholarlevelGraduateen_US
dc.identifier.bibliographicCitationB. Ahmed, T. A. Gulliver, S. alZahir, “Image splicing detection using mask-RCNN,” Signal, Image and Video Processing. pp. 1035-1042 (2020)en_US
dc.identifier.bibliographicCitationB. Ahmed, T. A. Gulliver, S. alZahir, “Blind copy-move forgery detection using SVD and KS test,” SN Applied Sciences. 2(8), pp. 1-2 (2020)en_US
dc.identifier.bibliographicCitationM. Reid, I. Hartley, E. Jensen, P. Kilcullen, B. Ahmed, A. Mohamed, and K. Lawyer, “THz NDE and applications for sensing and imaging of wood products,” Invited departmental seminar, Graduate Agricultural School, Nagoya University, Japan (Nov. 4, 2016).en_US
dc.identifier.bibliographicCitationM. Reid, B. Ahmed, I. Hartley, S. Tsuchikawa, and M. Reid, “Simultaneous prediction of density and moisture content of word by terahertz time domain spectroscopy,” Journal of Infrared, Millimeter, and Terahertz Waves, 35(11), pp. 949-961 (2014)en_US
dc.identifier.bibliographicCitationM. Reid, T. Inagaki, S. Tsuchikawa, K. Lawyer, B. Ahmed, A. Nasr, E.T. Jensen, and I.D. Hartley, “Terahertz spectroscopy of wood and combustion gas,” Invited presentation at the Canadian Association of Physics Congress, Sudbury, ON, CAN (June 18, 2014)en_US
dc.identifier.urihttp://hdl.handle.net/1828/12513
dc.languageEnglisheng
dc.language.isoenen_US
dc.rightsAvailable to the World Wide Weben_US
dc.subjectImage forgeryen_US
dc.subjectcopy-move forgeryen_US
dc.subjectimage splicing forgeryen_US
dc.subjectPADNETen_US
dc.subjectDeepUNeten_US
dc.subjectCoMoFoDen_US
dc.subjectsingular vector decomposition (SVD)en_US
dc.subjectsteerable pyramiden_US
dc.subjectMask-RCNNen_US
dc.titleDetection and localization of forgeries in digital imagesen_US
dc.typeThesisen_US

Files

Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
Ahmed_Belal_PhD_2020.pdf
Size:
87.09 MB
Format:
Adobe Portable Document Format
Description:
PhD Dissertation
Download
License bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
license.txt
Size:
1.71 KB
Format:
Item-specific license agreed upon to submission
Description:
Download

Collections

Electronic Theses and Dissertations (ETD)
Theses (Electrical and Computer Engineering)
 
University of Victoria Libraries

We acknowledge and respect the Lək̓ʷəŋən (Songhees and Esquimalt) Peoples on whose territory the university stands, and the Lək̓ʷəŋən and W̱SÁNEĆ Peoples whose historical relationships with the land continue to this day.