University of Twente Student Theses


Morphed Image Detection using Local Spectrum Analysis

Ndeh de Mbah, Eitel Yvan (2021) Morphed Image Detection using Local Spectrum Analysis.

[img] PDF
Abstract:According to research publications, face recognition systems are prone to morphed images attack. To address this issue, many Morphing Detection Attack techniques have been developed. These techniques are classified in three different categories based on the different images features exploited in the technique. One of the biggest challenges encountered when evaluating the performance of a developed technique is the fact that the technique is dependent on the morphed images dataset that was used in developing the technique. This morphing dataset is based on the method used to create the morphed images. This implies the developed technique may perform well on the database that was used to develop the technique, but once a different database is used, the developed method may no longer be able to detect morphed images accurately. In this work, a morphing attack detection (MAD) technique based on local spectrum analysis is presented. In this technique the absolute total power content over the spectrum of the images is used to differentiate between morphs and genuine images (mostly referred to as bona fides in biometrics). The analysis is made locally by splitting each facial images into six facial features and the classification based on each feature is attributed an individual binary score. This score refers to a right or wrong classification. To detect if a given image is a morphed image, these scores are then summed to obtain a total score and based on this score the proposed algorithm can determine if the image is morphed or not. FRGC database is used to train, validate and test the performance of the algorithm. A second database, an AMSL database is used to investigate the robustness of the proposed algorithm. Experimental analysis showed an ACER of 1.59% when the technique was tested on the FRGC database and an ACER of 38.24% on the AMSL database. ACER represents an error rate that was used to evaluate the performance of the technique on moprhed and genuine images.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:53 electrotechnology
Programme:Electrical Engineering BSc (56953)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page